slate-memory 0.1.0__tar.gz

slate_memory-0.1.0/LICENSE

@@ -0,0 +1,189 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work.
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to the Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by the Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding any notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   Copyright 2026 Matthew Lancaster
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

slate_memory-0.1.0/PKG-INFO

@@ -0,0 +1,161 @@
+Metadata-Version: 2.4
+Name: slate-memory
+Version: 0.1.0
+Summary: One-shot attractor memory for LLMs — vector-search accuracy at 98% fewer tokens.
+Author-email: Matthew Lancaster <mattclancaster@gmail.com>
+License-Expression: Apache-2.0
+Project-URL: Homepage, https://github.com/Scriblio/slate-memory
+Project-URL: Documentation, https://github.com/Scriblio/slate-memory#readme
+Keywords: llm,memory,rag,attractor,hopfield,retrieval
+Classifier: Development Status :: 4 - Beta
+Classifier: Intended Audience :: Developers
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Classifier: Programming Language :: Python :: 3
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: numpy>=1.24
+Provides-Extra: embed
+Requires-Dist: sentence-transformers>=2.0; extra == "embed"
+Dynamic: license-file
+
+# slate-memory
+
+One-shot attractor memory for LLMs. Vector-search accuracy at 98% fewer tokens.
+
+## The number
+
+| Setup | Accuracy | Cost per 1k queries | Latency |
+|---|---|---|---|
+| **haiku + slate** | **100%** | **$0.10** | **0.6s** |
+| opus + full context | 100% | $64.04 | 1.7s |
+| opus bare | 0% | $0.19 | 1.2s |
+| haiku bare | 0% | $0.04 | 0.7s |
+
+A cheap model with slate-memory **ties the most expensive model with the entire corpus in context** — at 1/640th the cost. Both models score 0% without memory (facts are synthetic, unknowable parametrically). Benchmark: N=25 questions, K=600 stored facts, hermetic SDK calls. Full methodology and reproduction scripts in [slate-bench](https://github.com/mattlancaster/slate-bench).
+
+## How it works
+
+Slate-memory is a [modern Hopfield network](https://arxiv.org/abs/2008.02217) (Ramsauer et al. 2020) — the same math as transformer attention, but used as a memory instead of a layer. Embeddings are sign-projected onto 10,000 bipolar cells (SimHash) and stored one-shot. On recall, the query settles into the nearest stored attractor via softmax-weighted feedback. Two cycles. No training. No index. No database.
+
+The attractor dynamics error-correct: noisy, partial, or corrupted queries converge to the exact stored pattern. Verified at 20,000 stored patterns with 25% corruption.
+
+## Install
+
+```bash
+pip install slate-memory
+```
+
+For the built-in embedder convenience wrapper:
+
+```bash
+pip install slate-memory[embed]
+```
+
+## Quick start
+
+```python
+from slate_memory import SlateBank
+import numpy as np
+
+# Your embeddings (384-dim for MiniLM, 1536 for OpenAI, etc.)
+bank = SlateBank(dim=384)
+
+# Commit facts one-shot
+bank.commit(embed("The capital of France is Paris"), {"text": "Paris is the capital of France", "id": "fact-1"})
+bank.commit(embed("Python was created by Guido van Rossum"), {"text": "Guido created Python", "id": "fact-2"})
+
+# Recall — even from a noisy or partial query
+winner, ranked, confidence, cycles = bank.recall(embed("What's the capital of France?"))
+print(winner)  # {"text": "Paris is the capital of France", "id": "fact-1"}
+print(f"Confidence: {confidence:.3f}, settled in {cycles} cycles")
+```
+
+### With sentence-transformers
+
+```python
+from slate_memory import SlateBank
+from sentence_transformers import SentenceTransformer
+
+model = SentenceTransformer("all-MiniLM-L6-v2")
+bank = SlateBank(dim=384)
+
+# Commit
+texts = ["The Earth orbits the Sun", "Water boils at 100°C", "Light travels at 299,792 km/s"]
+for text in texts:
+    bank.commit(model.encode(text), {"text": text})
+
+# Recall
+query = model.encode("what temperature does water boil")
+winner, ranked, confidence, cycles = bank.recall(query)
+print(winner["text"])  # "Water boils at 100°C"
+```
+
+### With OpenAI embeddings
+
+```python
+from slate_memory import SlateBank
+from openai import OpenAI
+
+client = OpenAI()
+bank = SlateBank(dim=1536)  # text-embedding-3-small
+
+def embed(text):
+    return client.embeddings.create(input=text, model="text-embedding-3-small").data[0].embedding
+
+bank.commit(embed("Revenue was $4.2M in Q3"), {"text": "Q3 revenue: $4.2M"})
+winner, _, conf, _ = bank.recall(embed("how much revenue in Q3"))
+```
+
+## Persistence
+
+```python
+# Save
+bank.save("./my_memory")
+
+# Load
+bank2 = SlateBank(dim=384)
+n = bank2.load("./my_memory")
+print(f"Loaded {n} patterns")
+```
+
+The entire memory is two files: `patterns.npy` (the attractor states) and `meta.json` (your metadata). No server. No database. Copy them anywhere.
+
+## API
+
+### `SlateBank(dim, n_cells=10000, beta=60.0, distinctiveness=True, dedup_threshold=0.95, seed=7)`
+
+- **dim**: Embedding dimensionality (must match your embedder)
+- **n_cells**: Bipolar cells in the attractor state. 10,000 is verified for up to 20k patterns
+- **beta**: Winner-take-all sharpness. 60.0 is benchmarked; don't change without reason
+- **distinctiveness**: Down-weight cells where all patterns agree. Essential for images; neutral for text
+- **dedup_threshold**: Refuse commits with overlap above this. 0.95 prevents near-duplicates
+- **seed**: RNG seed for the projection matrix. Same seed = same projections = portable patterns
+
+### `bank.commit(embedding, meta) → (bool, str)`
+
+One-shot storage. Returns `(True, "committed")` or `(False, "duplicate (...)")`.
+
+### `bank.recall(embedding, top_k=3, max_cycles=5) → (winner, ranked, confidence, cycles)`
+
+Full attractor settle. Returns the winner's metadata, top-k ranked results, confidence score, and settle cycles.
+
+### `bank.familiar(embedding) → (meta, score)`
+
+Quick overlap check without settling. Use for "have I seen this before?" checks.
+
+### `bank.save(path)` / `bank.load(path)`
+
+Persist to / restore from a directory.
+
+## Benchmarks
+
+Full retrieval benchmarks across 5 corruption conditions and 4 capacity levels show accuracy parity with exact cosine search (±1 point). The attractor substrate recalls 20,000 random patterns perfectly at 25% corruption. Capacity is limited by embedding discriminability (the embedder), not the attractor.
+
+Honest limitations in software: per-query latency is ~35x slower than exact vector search (19ms vs 0.5ms at K=2000) because the projected patterns are 10,000-d vs 384-d. The speed case is the [photonic implementation](https://en.wikipedia.org/wiki/Optical_computing) where recall is one pass of light — that's the roadmap, not the demo.
+
+See [slate-bench](https://github.com/mattlancaster/slate-bench) for full reproduction scripts and results.
+
+## License
+
+Apache 2.0. Patent pending.

slate_memory-0.1.0/README.md

@@ -0,0 +1,140 @@
+# slate-memory
+
+One-shot attractor memory for LLMs. Vector-search accuracy at 98% fewer tokens.
+
+## The number
+
+| Setup | Accuracy | Cost per 1k queries | Latency |
+|---|---|---|---|
+| **haiku + slate** | **100%** | **$0.10** | **0.6s** |
+| opus + full context | 100% | $64.04 | 1.7s |
+| opus bare | 0% | $0.19 | 1.2s |
+| haiku bare | 0% | $0.04 | 0.7s |
+
+A cheap model with slate-memory **ties the most expensive model with the entire corpus in context** — at 1/640th the cost. Both models score 0% without memory (facts are synthetic, unknowable parametrically). Benchmark: N=25 questions, K=600 stored facts, hermetic SDK calls. Full methodology and reproduction scripts in [slate-bench](https://github.com/mattlancaster/slate-bench).
+
+## How it works
+
+Slate-memory is a [modern Hopfield network](https://arxiv.org/abs/2008.02217) (Ramsauer et al. 2020) — the same math as transformer attention, but used as a memory instead of a layer. Embeddings are sign-projected onto 10,000 bipolar cells (SimHash) and stored one-shot. On recall, the query settles into the nearest stored attractor via softmax-weighted feedback. Two cycles. No training. No index. No database.
+
+The attractor dynamics error-correct: noisy, partial, or corrupted queries converge to the exact stored pattern. Verified at 20,000 stored patterns with 25% corruption.
+
+## Install
+
+```bash
+pip install slate-memory
+```
+
+For the built-in embedder convenience wrapper:
+
+```bash
+pip install slate-memory[embed]
+```
+
+## Quick start
+
+```python
+from slate_memory import SlateBank
+import numpy as np
+
+# Your embeddings (384-dim for MiniLM, 1536 for OpenAI, etc.)
+bank = SlateBank(dim=384)
+
+# Commit facts one-shot
+bank.commit(embed("The capital of France is Paris"), {"text": "Paris is the capital of France", "id": "fact-1"})
+bank.commit(embed("Python was created by Guido van Rossum"), {"text": "Guido created Python", "id": "fact-2"})
+
+# Recall — even from a noisy or partial query
+winner, ranked, confidence, cycles = bank.recall(embed("What's the capital of France?"))
+print(winner)  # {"text": "Paris is the capital of France", "id": "fact-1"}
+print(f"Confidence: {confidence:.3f}, settled in {cycles} cycles")
+```
+
+### With sentence-transformers
+
+```python
+from slate_memory import SlateBank
+from sentence_transformers import SentenceTransformer
+
+model = SentenceTransformer("all-MiniLM-L6-v2")
+bank = SlateBank(dim=384)
+
+# Commit
+texts = ["The Earth orbits the Sun", "Water boils at 100°C", "Light travels at 299,792 km/s"]
+for text in texts:
+    bank.commit(model.encode(text), {"text": text})
+
+# Recall
+query = model.encode("what temperature does water boil")
+winner, ranked, confidence, cycles = bank.recall(query)
+print(winner["text"])  # "Water boils at 100°C"
+```
+
+### With OpenAI embeddings
+
+```python
+from slate_memory import SlateBank
+from openai import OpenAI
+
+client = OpenAI()
+bank = SlateBank(dim=1536)  # text-embedding-3-small
+
+def embed(text):
+    return client.embeddings.create(input=text, model="text-embedding-3-small").data[0].embedding
+
+bank.commit(embed("Revenue was $4.2M in Q3"), {"text": "Q3 revenue: $4.2M"})
+winner, _, conf, _ = bank.recall(embed("how much revenue in Q3"))
+```
+
+## Persistence
+
+```python
+# Save
+bank.save("./my_memory")
+
+# Load
+bank2 = SlateBank(dim=384)
+n = bank2.load("./my_memory")
+print(f"Loaded {n} patterns")
+```
+
+The entire memory is two files: `patterns.npy` (the attractor states) and `meta.json` (your metadata). No server. No database. Copy them anywhere.
+
+## API
+
+### `SlateBank(dim, n_cells=10000, beta=60.0, distinctiveness=True, dedup_threshold=0.95, seed=7)`
+
+- **dim**: Embedding dimensionality (must match your embedder)
+- **n_cells**: Bipolar cells in the attractor state. 10,000 is verified for up to 20k patterns
+- **beta**: Winner-take-all sharpness. 60.0 is benchmarked; don't change without reason
+- **distinctiveness**: Down-weight cells where all patterns agree. Essential for images; neutral for text
+- **dedup_threshold**: Refuse commits with overlap above this. 0.95 prevents near-duplicates
+- **seed**: RNG seed for the projection matrix. Same seed = same projections = portable patterns
+
+### `bank.commit(embedding, meta) → (bool, str)`
+
+One-shot storage. Returns `(True, "committed")` or `(False, "duplicate (...)")`.
+
+### `bank.recall(embedding, top_k=3, max_cycles=5) → (winner, ranked, confidence, cycles)`
+
+Full attractor settle. Returns the winner's metadata, top-k ranked results, confidence score, and settle cycles.
+
+### `bank.familiar(embedding) → (meta, score)`
+
+Quick overlap check without settling. Use for "have I seen this before?" checks.
+
+### `bank.save(path)` / `bank.load(path)`
+
+Persist to / restore from a directory.
+
+## Benchmarks
+
+Full retrieval benchmarks across 5 corruption conditions and 4 capacity levels show accuracy parity with exact cosine search (±1 point). The attractor substrate recalls 20,000 random patterns perfectly at 25% corruption. Capacity is limited by embedding discriminability (the embedder), not the attractor.
+
+Honest limitations in software: per-query latency is ~35x slower than exact vector search (19ms vs 0.5ms at K=2000) because the projected patterns are 10,000-d vs 384-d. The speed case is the [photonic implementation](https://en.wikipedia.org/wiki/Optical_computing) where recall is one pass of light — that's the roadmap, not the demo.
+
+See [slate-bench](https://github.com/mattlancaster/slate-bench) for full reproduction scripts and results.
+
+## License
+
+Apache 2.0. Patent pending.

slate_memory-0.1.0/pyproject.toml

@@ -0,0 +1,34 @@
+[build-system]
+requires = ["setuptools>=68.0", "wheel"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "slate-memory"
+version = "0.1.0"
+description = "One-shot attractor memory for LLMs — vector-search accuracy at 98% fewer tokens."
+readme = "README.md"
+license = "Apache-2.0"
+requires-python = ">=3.10"
+authors = [
+    {name = "Matthew Lancaster", email = "mattclancaster@gmail.com"},
+]
+keywords = ["llm", "memory", "rag", "attractor", "hopfield", "retrieval"]
+classifiers = [
+    "Development Status :: 4 - Beta",
+    "Intended Audience :: Developers",
+    "Topic :: Scientific/Engineering :: Artificial Intelligence",
+    "Programming Language :: Python :: 3",
+]
+dependencies = [
+    "numpy>=1.24",
+]
+
+[project.optional-dependencies]
+embed = ["sentence-transformers>=2.0"]
+
+[project.urls]
+Homepage = "https://github.com/Scriblio/slate-memory"
+Documentation = "https://github.com/Scriblio/slate-memory#readme"
+
+[tool.setuptools.packages.find]
+include = ["slate_memory*"]

slate_memory-0.1.0/setup.cfg

@@ -0,0 +1,4 @@
+[egg_info]
+tag_build = 
+tag_date = 0
+

slate_memory-0.1.0/slate_memory/__init__.py

@@ -0,0 +1,13 @@
+"""slate-memory — one-shot attractor memory for LLMs and agents.
+
+    from slate_memory import SlateBank
+
+    bank = SlateBank()
+    bank.commit(embedding, {"text": "Paris is the capital of France"})
+    winner, top_k, confidence, cycles = bank.recall(query_embedding)
+"""
+
+from slate_memory.bank import SlateBank
+
+__all__ = ["SlateBank"]
+__version__ = "0.1.0"

slate_memory-0.1.0/slate_memory/bank.py

@@ -0,0 +1,244 @@
+"""SlateBank — one-shot attractor memory with error-correcting recall.
+
+Text or image embeddings are sign-projected onto N bipolar cells (SimHash),
+converting cosine similarity into Hamming similarity, and committed one-shot.
+Recall settles the probe into the nearest stored attractor via a softmax-
+weighted (modern Hopfield) feedback update.
+
+Benchmarked: accuracy parity with exact vector search; 98% prompt-token
+savings over context stuffing; haiku+slate ties opus+full-context at 1/640th
+the cost.
+"""
+
+from __future__ import annotations
+
+import json
+import threading
+from pathlib import Path
+
+import numpy as np
+
+
+class SlateBank:
+    """One-shot attractor memory with error-correcting recall.
+
+    Parameters
+    ----------
+    n_cells : int
+        Number of bipolar cells in the attractor state. More cells = more
+        capacity, but slower recall. 10,000 is the verified sweet spot.
+    dim : int
+        Dimensionality of input embeddings. Must match your embedder
+        (384 for MiniLM, 768 for many others, 1536 for OpenAI).
+    beta : float
+        Winner-take-all sharpness. Higher = harder selection among stored
+        patterns. 60.0 is verified; don't change without benchmarking.
+    distinctiveness : bool
+        Down-weight cells where all stored patterns agree (shared template
+        scaffolding). Essential for visual patterns; neutral for text.
+    dedup_threshold : float
+        Overlap above which a new commit is refused as a duplicate.
+        0.95 is conservative; lower to allow finer distinctions.
+    seed : int
+        RNG seed for the projection matrix. Two SlateBank instances with
+        the same seed + dim + n_cells produce identical projections, so
+        patterns are portable between them.
+    """
+
+    def __init__(
+        self,
+        n_cells: int = 10_000,
+        dim: int = 384,
+        beta: float = 60.0,
+        distinctiveness: bool = True,
+        dedup_threshold: float = 0.95,
+        seed: int = 7,
+    ):
+        rng = np.random.default_rng(seed)
+        self._R = rng.standard_normal((n_cells, dim)).astype(np.float32)
+        self._n = n_cells
+        self._beta = beta
+        self._distinctiveness = distinctiveness
+        self._dedup_threshold = dedup_threshold
+        self._lock = threading.Lock()
+        self._patterns: list[np.ndarray] = []
+        self._meta: list[dict] = []
+        self._P: np.ndarray | None = None
+        self._w: np.ndarray | None = None
+
+    @property
+    def count(self) -> int:
+        """Number of stored patterns."""
+        return len(self._patterns)
+
+    def _project(self, emb: np.ndarray) -> np.ndarray:
+        e = np.asarray(emb, dtype=np.float32).ravel()
+        return np.where(self._R @ e >= 0, 1.0, -1.0).astype(np.float32)
+
+    def _finalize(self) -> None:
+        if not self._patterns:
+            return
+        self._P = np.stack(self._patterns)
+        if self._distinctiveness and len(self._patterns) >= 2:
+            agree = np.abs(self._P.mean(axis=0))
+            w = 1.0 - agree
+            s = float(w.sum())
+            if s > self._n * 0.05:
+                self._w = (w * (self._n / s)).astype(np.float32)
+            else:
+                self._w = None
+        else:
+            self._w = None
+
+    def _overlaps(self, s: np.ndarray) -> np.ndarray:
+        if self._w is not None:
+            return (self._P @ (self._w * s)) / self._n
+        return (self._P @ s) / self._n
+
+    # ── write ────────────────────────────────────────────────────────────
+
+    def commit(self, embedding: np.ndarray, meta: dict) -> tuple[bool, str]:
+        """Store a pattern one-shot. Returns (committed, reason).
+
+        Parameters
+        ----------
+        embedding : array-like
+            The embedding vector (must match `dim`).
+        meta : dict
+            Arbitrary metadata returned on recall (text, id, tags, etc.).
+        """
+        pat = self._project(embedding)
+        with self._lock:
+            if self._patterns and self._dedup_threshold < 1.0:
+                if self._P is None:
+                    self._finalize()
+                overlaps = (self._P @ pat) / self._n
+                best = float(np.max(overlaps))
+                if best >= self._dedup_threshold:
+                    idx = int(np.argmax(overlaps))
+                    return False, f"duplicate (overlap {best:.3f} with {self._meta[idx].get('id', idx)})"
+            self._patterns.append(pat)
+            self._meta.append(meta)
+            self._P = None
+            return True, "committed"
+
+    def commit_batch(
+        self, embeddings: list[np.ndarray], metas: list[dict]
+    ) -> list[tuple[bool, str]]:
+        """Commit multiple patterns. Dedup checks run against the growing store."""
+        results = []
+        for emb, meta in zip(embeddings, metas):
+            results.append(self.commit(emb, meta))
+        return results
+
+    # ── read ─────────────────────────────────────────────────────────────
+
+    def recall(
+        self, embedding: np.ndarray, top_k: int = 3, max_cycles: int = 5
+    ) -> tuple[dict | None, list[dict], float, int]:
+        """Settle into the nearest attractor. Returns (winner, ranked, confidence, cycles).
+
+        Parameters
+        ----------
+        embedding : array-like
+            Query embedding.
+        top_k : int
+            Number of ranked results to return.
+        max_cycles : int
+            Maximum settle iterations. Usually converges in 2.
+        """
+        with self._lock:
+            if not self._patterns:
+                return None, [], 0.0, 0
+            if self._P is None:
+                self._finalize()
+            s = self._project(embedding)
+            initial = self._overlaps(s)
+            cycles = 0
+            for _ in range(max_cycles):
+                o = self._overlaps(s)
+                a = np.exp(self._beta * (o - o.max()))
+                a /= a.sum()
+                s_new = np.where(self._P.T @ a >= 0, 1.0, -1.0).astype(np.float32)
+                cycles += 1
+                if np.array_equal(s_new, s):
+                    break
+                s = s_new
+            final = self._overlaps(s)
+            winner = int(np.argmax(final))
+            order = np.argsort(-initial)[:top_k]
+            ranked = [self._meta[int(i)] for i in order]
+            if ranked and ranked[0] != self._meta[winner]:
+                ranked = [self._meta[winner]] + [
+                    m for m in ranked if m != self._meta[winner]
+                ][:top_k - 1]
+            return self._meta[winner], ranked, float(final[winner]), cycles
+
+    def familiar(self, embedding: np.ndarray) -> tuple[dict | None, float]:
+        """Quick similarity check without full settle. Returns (best_meta, score)."""
+        with self._lock:
+            if not self._patterns:
+                return None, 0.0
+            if self._P is None:
+                self._finalize()
+            s = self._project(embedding)
+            o = self._overlaps(s)
+            best = int(np.argmax(o))
+            return self._meta[best], float(o[best])
+
+    # ── persistence ──────────────────────────────────────────────────────
+
+    def save(self, path: str | Path) -> None:
+        """Save patterns and metadata to a directory."""
+        d = Path(path)
+        d.mkdir(parents=True, exist_ok=True)
+        with self._lock:
+            if self._patterns:
+                P = np.stack(self._patterns)
+                np.save(str(d / "patterns.npy"), P)
+            with open(d / "meta.json", "w", encoding="utf-8") as f:
+                json.dump(self._meta, f, ensure_ascii=False, indent=1)
+            with open(d / "config.json", "w") as f:
+                json.dump({
+                    "n_cells": self._n,
+                    "dim": self._R.shape[1],
+                    "beta": self._beta,
+                    "seed": 7,
+                    "distinctiveness": self._distinctiveness,
+                    "dedup_threshold": self._dedup_threshold,
+                    "count": len(self._patterns),
+                }, f, indent=2)
+
+    def load(self, path: str | Path) -> int:
+        """Load patterns and metadata. Returns number of patterns loaded."""
+        d = Path(path)
+        if not (d / "patterns.npy").exists():
+            return 0
+        with self._lock:
+            P = np.load(str(d / "patterns.npy"))
+            with open(d / "meta.json", encoding="utf-8") as f:
+                self._meta = json.load(f)
+            n = min(len(P), len(self._meta))
+            self._patterns = [P[i] for i in range(n)]
+            self._meta = self._meta[:n]
+            self._P = None
+            return n
+
+    # ── utility ──────────────────────────────────────────────────────────
+
+    def list(self, limit: int = 50) -> list[dict]:
+        """Return metadata for stored patterns."""
+        with self._lock:
+            return list(self._meta[:limit])
+
+    def stats(self) -> dict:
+        """Return summary statistics."""
+        with self._lock:
+            return {
+                "count": len(self._patterns),
+                "n_cells": self._n,
+                "dim": self._R.shape[1],
+                "beta": self._beta,
+                "distinctiveness": self._distinctiveness,
+                "dedup_threshold": self._dedup_threshold,
+            }

slate_memory-0.1.0/slate_memory.egg-info/PKG-INFO

@@ -0,0 +1,161 @@
+Metadata-Version: 2.4
+Name: slate-memory
+Version: 0.1.0
+Summary: One-shot attractor memory for LLMs — vector-search accuracy at 98% fewer tokens.
+Author-email: Matthew Lancaster <mattclancaster@gmail.com>
+License-Expression: Apache-2.0
+Project-URL: Homepage, https://github.com/Scriblio/slate-memory
+Project-URL: Documentation, https://github.com/Scriblio/slate-memory#readme
+Keywords: llm,memory,rag,attractor,hopfield,retrieval
+Classifier: Development Status :: 4 - Beta
+Classifier: Intended Audience :: Developers
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Classifier: Programming Language :: Python :: 3
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: numpy>=1.24
+Provides-Extra: embed
+Requires-Dist: sentence-transformers>=2.0; extra == "embed"
+Dynamic: license-file
+
+# slate-memory
+
+One-shot attractor memory for LLMs. Vector-search accuracy at 98% fewer tokens.
+
+## The number
+
+| Setup | Accuracy | Cost per 1k queries | Latency |
+|---|---|---|---|
+| **haiku + slate** | **100%** | **$0.10** | **0.6s** |
+| opus + full context | 100% | $64.04 | 1.7s |
+| opus bare | 0% | $0.19 | 1.2s |
+| haiku bare | 0% | $0.04 | 0.7s |
+
+A cheap model with slate-memory **ties the most expensive model with the entire corpus in context** — at 1/640th the cost. Both models score 0% without memory (facts are synthetic, unknowable parametrically). Benchmark: N=25 questions, K=600 stored facts, hermetic SDK calls. Full methodology and reproduction scripts in [slate-bench](https://github.com/mattlancaster/slate-bench).
+
+## How it works
+
+Slate-memory is a [modern Hopfield network](https://arxiv.org/abs/2008.02217) (Ramsauer et al. 2020) — the same math as transformer attention, but used as a memory instead of a layer. Embeddings are sign-projected onto 10,000 bipolar cells (SimHash) and stored one-shot. On recall, the query settles into the nearest stored attractor via softmax-weighted feedback. Two cycles. No training. No index. No database.
+
+The attractor dynamics error-correct: noisy, partial, or corrupted queries converge to the exact stored pattern. Verified at 20,000 stored patterns with 25% corruption.
+
+## Install
+
+```bash
+pip install slate-memory
+```
+
+For the built-in embedder convenience wrapper:
+
+```bash
+pip install slate-memory[embed]
+```
+
+## Quick start
+
+```python
+from slate_memory import SlateBank
+import numpy as np
+
+# Your embeddings (384-dim for MiniLM, 1536 for OpenAI, etc.)
+bank = SlateBank(dim=384)
+
+# Commit facts one-shot
+bank.commit(embed("The capital of France is Paris"), {"text": "Paris is the capital of France", "id": "fact-1"})
+bank.commit(embed("Python was created by Guido van Rossum"), {"text": "Guido created Python", "id": "fact-2"})
+
+# Recall — even from a noisy or partial query
+winner, ranked, confidence, cycles = bank.recall(embed("What's the capital of France?"))
+print(winner)  # {"text": "Paris is the capital of France", "id": "fact-1"}
+print(f"Confidence: {confidence:.3f}, settled in {cycles} cycles")
+```
+
+### With sentence-transformers
+
+```python
+from slate_memory import SlateBank
+from sentence_transformers import SentenceTransformer
+
+model = SentenceTransformer("all-MiniLM-L6-v2")
+bank = SlateBank(dim=384)
+
+# Commit
+texts = ["The Earth orbits the Sun", "Water boils at 100°C", "Light travels at 299,792 km/s"]
+for text in texts:
+    bank.commit(model.encode(text), {"text": text})
+
+# Recall
+query = model.encode("what temperature does water boil")
+winner, ranked, confidence, cycles = bank.recall(query)
+print(winner["text"])  # "Water boils at 100°C"
+```
+
+### With OpenAI embeddings
+
+```python
+from slate_memory import SlateBank
+from openai import OpenAI
+
+client = OpenAI()
+bank = SlateBank(dim=1536)  # text-embedding-3-small
+
+def embed(text):
+    return client.embeddings.create(input=text, model="text-embedding-3-small").data[0].embedding
+
+bank.commit(embed("Revenue was $4.2M in Q3"), {"text": "Q3 revenue: $4.2M"})
+winner, _, conf, _ = bank.recall(embed("how much revenue in Q3"))
+```
+
+## Persistence
+
+```python
+# Save
+bank.save("./my_memory")
+
+# Load
+bank2 = SlateBank(dim=384)
+n = bank2.load("./my_memory")
+print(f"Loaded {n} patterns")
+```
+
+The entire memory is two files: `patterns.npy` (the attractor states) and `meta.json` (your metadata). No server. No database. Copy them anywhere.
+
+## API
+
+### `SlateBank(dim, n_cells=10000, beta=60.0, distinctiveness=True, dedup_threshold=0.95, seed=7)`
+
+- **dim**: Embedding dimensionality (must match your embedder)
+- **n_cells**: Bipolar cells in the attractor state. 10,000 is verified for up to 20k patterns
+- **beta**: Winner-take-all sharpness. 60.0 is benchmarked; don't change without reason
+- **distinctiveness**: Down-weight cells where all patterns agree. Essential for images; neutral for text
+- **dedup_threshold**: Refuse commits with overlap above this. 0.95 prevents near-duplicates
+- **seed**: RNG seed for the projection matrix. Same seed = same projections = portable patterns
+
+### `bank.commit(embedding, meta) → (bool, str)`
+
+One-shot storage. Returns `(True, "committed")` or `(False, "duplicate (...)")`.
+
+### `bank.recall(embedding, top_k=3, max_cycles=5) → (winner, ranked, confidence, cycles)`
+
+Full attractor settle. Returns the winner's metadata, top-k ranked results, confidence score, and settle cycles.
+
+### `bank.familiar(embedding) → (meta, score)`
+
+Quick overlap check without settling. Use for "have I seen this before?" checks.
+
+### `bank.save(path)` / `bank.load(path)`
+
+Persist to / restore from a directory.
+
+## Benchmarks
+
+Full retrieval benchmarks across 5 corruption conditions and 4 capacity levels show accuracy parity with exact cosine search (±1 point). The attractor substrate recalls 20,000 random patterns perfectly at 25% corruption. Capacity is limited by embedding discriminability (the embedder), not the attractor.
+
+Honest limitations in software: per-query latency is ~35x slower than exact vector search (19ms vs 0.5ms at K=2000) because the projected patterns are 10,000-d vs 384-d. The speed case is the [photonic implementation](https://en.wikipedia.org/wiki/Optical_computing) where recall is one pass of light — that's the roadmap, not the demo.
+
+See [slate-bench](https://github.com/mattlancaster/slate-bench) for full reproduction scripts and results.
+
+## License
+
+Apache 2.0. Patent pending.

slate_memory-0.1.0/slate_memory.egg-info/SOURCES.txt

@@ -0,0 +1,11 @@
+LICENSE
+README.md
+pyproject.toml
+slate_memory/__init__.py
+slate_memory/bank.py
+slate_memory.egg-info/PKG-INFO
+slate_memory.egg-info/SOURCES.txt
+slate_memory.egg-info/dependency_links.txt
+slate_memory.egg-info/requires.txt
+slate_memory.egg-info/top_level.txt
+tests/test_bank.py

slate_memory-0.1.0/slate_memory.egg-info/dependency_links.txt

@@ -0,0 +1 @@
+

slate_memory-0.1.0/slate_memory.egg-info/requires.txt

@@ -0,0 +1,4 @@
+numpy>=1.24
+
+[embed]
+sentence-transformers>=2.0

slate_memory-0.1.0/slate_memory.egg-info/top_level.txt

@@ -0,0 +1 @@
+slate_memory

slate_memory-0.1.0/tests/test_bank.py

@@ -0,0 +1,140 @@
+"""Tests for SlateBank — the core attractor memory."""
+
+import tempfile
+from pathlib import Path
+
+import numpy as np
+import pytest
+
+from slate_memory import SlateBank
+
+
+@pytest.fixture
+def bank():
+    return SlateBank(n_cells=1000, dim=64, seed=42)
+
+
+@pytest.fixture
+def rng():
+    return np.random.default_rng(99)
+
+
+def _random_emb(rng, dim=64):
+    e = rng.standard_normal(dim).astype(np.float32)
+    return e / np.linalg.norm(e)
+
+
+def test_commit_and_recall(bank, rng):
+    emb = _random_emb(rng)
+    ok, reason = bank.commit(emb, {"id": "a"})
+    assert ok
+    assert reason == "committed"
+    assert bank.count == 1
+
+    winner, ranked, conf, cycles = bank.recall(emb)
+    assert winner["id"] == "a"
+    assert conf > 0.9
+
+
+def test_recall_empty(bank, rng):
+    winner, ranked, conf, cycles = bank.recall(_random_emb(rng))
+    assert winner is None
+    assert ranked == []
+
+
+def test_noisy_recall(bank, rng):
+    emb = _random_emb(rng)
+    bank.commit(emb, {"id": "original"})
+    noisy = emb + rng.standard_normal(64).astype(np.float32) * 0.3
+    winner, _, conf, _ = bank.recall(noisy)
+    assert winner["id"] == "original"
+
+
+def test_multiple_patterns(bank, rng):
+    embs = [_random_emb(rng) for _ in range(20)]
+    for i, e in enumerate(embs):
+        bank.commit(e, {"id": f"fact-{i}"})
+    assert bank.count == 20
+    for i, e in enumerate(embs):
+        winner, _, _, _ = bank.recall(e)
+        assert winner["id"] == f"fact-{i}"
+
+
+def test_dedup_guard(bank, rng):
+    emb = _random_emb(rng)
+    ok1, _ = bank.commit(emb, {"id": "first"})
+    ok2, reason = bank.commit(emb, {"id": "second"})
+    assert ok1
+    assert not ok2
+    assert "duplicate" in reason
+    assert bank.count == 1
+
+
+def test_familiar(bank, rng):
+    emb = _random_emb(rng)
+    meta, score = bank.familiar(emb)
+    assert meta is None
+    assert score == 0.0
+
+    bank.commit(emb, {"id": "x"})
+    meta, score = bank.familiar(emb)
+    assert meta["id"] == "x"
+    assert score > 0.9
+
+
+def test_save_load(bank, rng):
+    for i in range(5):
+        bank.commit(_random_emb(rng), {"id": f"p-{i}"})
+
+    with tempfile.TemporaryDirectory() as d:
+        bank.save(d)
+        assert (Path(d) / "patterns.npy").exists()
+        assert (Path(d) / "meta.json").exists()
+
+        bank2 = SlateBank(n_cells=1000, dim=64, seed=42)
+        n = bank2.load(d)
+        assert n == 5
+        assert bank2.count == 5
+
+
+def test_persistence_recall_matches(bank, rng):
+    embs = [_random_emb(rng) for _ in range(10)]
+    for i, e in enumerate(embs):
+        bank.commit(e, {"id": f"r-{i}"})
+
+    with tempfile.TemporaryDirectory() as d:
+        bank.save(d)
+        bank2 = SlateBank(n_cells=1000, dim=64, seed=42)
+        bank2.load(d)
+        for i, e in enumerate(embs):
+            w1, _, c1, _ = bank.recall(e)
+            w2, _, c2, _ = bank2.recall(e)
+            assert w1["id"] == w2["id"]
+
+
+def test_commit_batch(bank, rng):
+    embs = [_random_emb(rng) for _ in range(5)]
+    metas = [{"id": f"b-{i}"} for i in range(5)]
+    results = bank.commit_batch(embs, metas)
+    assert all(ok for ok, _ in results)
+    assert bank.count == 5
+
+
+def test_stats(bank, rng):
+    bank.commit(_random_emb(rng), {"id": "s"})
+    s = bank.stats()
+    assert s["count"] == 1
+    assert s["n_cells"] == 1000
+    assert s["dim"] == 64
+
+
+def test_different_dims():
+    for dim in [384, 768, 1536]:
+        b = SlateBank(dim=dim, n_cells=500)
+        rng = np.random.default_rng(0)
+        e = rng.standard_normal(dim).astype(np.float32)
+        ok, _ = b.commit(e, {"dim": dim})
+        assert ok
+        w, _, c, _ = b.recall(e)
+        assert w["dim"] == dim
+        assert c > 0.8