engram-lite 0.1.0__tar.gz

engram_lite-0.1.0/.gitignore

@@ -0,0 +1,37 @@
+# python
+__pycache__/
+*.py[cod]
+*.egg-info/
+dist/
+build/
+.venv/
+venv/
+
+# tooling
+.pytest_cache/
+.ruff_cache/
+.coverage
+
+# local memory stores (never commit a memory DB)
+*.db
+*.db-wal
+*.db-shm
+*.sqlite
+*.sqlite3
+
+# model cache (fastembed downloads here)
+.fastembed_cache/
+.cache/
+
+# PyInstaller build outputs
+*.spec.bak
+
+# os
+.DS_Store
+
+# benchmark data (CC BY-NC — never redistributed) + local artifacts
+benchmarks/locomo/data/
+benchmarks/locomo/locomo/
+benchmarks/locomo/*.jsonl
+.env
+*.log

engram_lite-0.1.0/CHANGELOG.md

@@ -0,0 +1,15 @@
+# Changelog
+
+## 0.1.0 — first public release
+
+- Local memory engine: SQLite + FTS5 + sqlite-vec in one file; salience gate,
+  ADD / UPDATE / REINFORCE consolidation, versioned facts with provenance,
+  expiry, and bounded size with eviction.
+- **Conditioned promotion (the lane model):** register a profile per agent
+  (persona / domain / scope_tags) and every recall is served for that agent's
+  lane — IDF-weighted, relevance-floored, with honest abstention and an
+  explainable `why` on every served fact. Measured on our internal CAMP-Bench:
+  profile discrimination 0.00 → 0.90, task success 40% → 80%, hallucination
+  cut to a third versus a flat shared pile.
+- Hermes integration: `engram.integrations.hermes.EngramMemoryProvider`
+  (initialize / system_prompt_block / prefetch / sync_turn / tools / shutdown).

engram_lite-0.1.0/LICENSE

@@ -0,0 +1,202 @@
+
+                                 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
+      (an example is provided in the Appendix below).
+
+      "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 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 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 those 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
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   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.

engram_lite-0.1.0/PKG-INFO

@@ -0,0 +1,199 @@
+Metadata-Version: 2.4
+Name: engram-lite
+Version: 0.1.0
+Summary: Local memory for AI agents — served by persona, domain, and task. Survives restarts. One machine, many agents.
+Project-URL: Homepage, https://github.com/engrammemory-labs/engram-lite
+Project-URL: Repository, https://github.com/engrammemory-labs/engram-lite
+Project-URL: Issues, https://github.com/engrammemory-labs/engram-lite/issues
+Project-URL: Changelog, https://github.com/engrammemory-labs/engram-lite/blob/main/CHANGELOG.md
+Author: Novarque
+License: Apache-2.0
+License-File: LICENSE
+Keywords: agents,ai,hermes,llm,memory
+Classifier: Development Status :: 4 - Beta
+Classifier: Intended Audience :: Developers
+Classifier: License :: OSI Approved :: Apache Software License
+Classifier: Operating System :: OS Independent
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Programming Language :: Python :: 3.13
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.10
+Requires-Dist: fastembed>=0.3.6
+Requires-Dist: numpy>=1.26
+Requires-Dist: sqlite-vec>=0.1.6
+Provides-Extra: build
+Requires-Dist: pyinstaller>=6; extra == 'build'
+Provides-Extra: dev
+Requires-Dist: anyio>=4; extra == 'dev'
+Requires-Dist: pytest>=8; extra == 'dev'
+Requires-Dist: ruff>=0.8; extra == 'dev'
+Description-Content-Type: text/markdown
+
+# engram-lite
+
+**Local memory for AI agents, served by persona, domain, and task. Survives restarts. One machine, many agents.**
+
+Most agent memory is a shared pile: every agent gets the same top-k for the same query. In our benchmark (methodology below), an agent fed flat shared memory **hallucinated more than an agent given no memory at all.** Serving each agent the memory that fits **who it is** (its persona, its domain, the task in front of it) **doubled task success (40% to 80%)** and **cut hallucination to a third.**
+
+And the memory layer never calls an LLM. On **LoCoMo** — the long-conversation memory benchmark — engram-lite scores **J 68.3** with **zero LLM calls and $0 to build the memory**. The protocol, footnotes, and the full comparison against the leading LLM-based memory systems live in [`benchmarks/locomo/`](https://github.com/engrammemory-labs/engram-lite/blob/main/benchmarks/locomo/README.md).
+
+engram-lite is that serving layer, small enough to run on your laptop:
+
+- **Conditioned serving.** Register a profile per agent (persona / domain / scope). The same question serves your on-call SRE agent the alert, the trace, and the mitigation runbook, and serves your release agent the deploy, the rollback, and the freeze policy. Out-of-lane questions correctly serve nothing.
+- **Restart-proof.** One SQLite file holds everything. Your agent restarts tomorrow and picks up exactly what it knew.
+- **Self-cleaning.** A salience gate skips junk (code, command output, questions). New facts are de-duplicated, updates supersede old versions, stale facts expire, and the store stays bounded.
+- **Zero infrastructure.** No server, no cloud, no accounts, no telemetry. SQLite plus a local embedding model. `pip install` and you are running.
+- **Explainable forgetting.** Every drop, merge, truncation, and abstention is recorded with the rule that fired (`Memory.decisions()`). "Why don't you remember X?" has an answer — something an LLM-written memory can't give you, because its keep/drop decisions live inside model weights.
+- **Deterministic and replayable.** Same transcript in, same memory out — byte-identical across rebuilds. Memory failures can be diffed, bisected, and regression-gated like any other bug.
+- **Plugs into what you build with.** A Hermes memory provider, or three lines of Python.
+
+> engram-lite is single-machine by design. A team/hosted edition, **engram-core**, is in development — reach out if that is what you need.
+
+## Install
+
+```bash
+pip install engram-lite
+```
+
+Or Docker (memory persists in the named volume):
+
+```bash
+docker build -t engram-lite .
+docker run -it --rm -v engram-data:/data engram-lite demo
+```
+
+## 60 seconds: see conditioned memory work
+
+```bash
+python examples/restart_demo.py    # two agents remember
+python examples/restart_demo.py    # run it AGAIN: memory survived, and the same
+                                   # question serves each agent different memory
+```
+
+## Use it from Python
+
+```python
+from engram import Memory
+
+mem = Memory("~/.engram/memory.db", origin_tool="oncall_sre")
+
+# who is this agent? (this is what makes serving conditioned)
+mem.register_profile(
+    "oncall_sre", persona="on-call SRE", domain="sre-devops",
+    scope_tags=["alert", "incident", "latency", "trace", "mitigation", "db"],
+)
+
+mem.remember("Alert PAGE-99 fired: payments p99 latency at 2.3s")   # gated save
+# conversational input? pass who said it and when — every extracted fragment
+# keeps the anchors, and relative dates resolve to absolute ("last Friday"
+# becomes "[= Friday, 5 May 2023]") with calendar math, no LLM:
+mem.remember("The pool was resized last Friday. We watch it daily now.",
+             speaker="Raj", when="8 May, 2023")
+hits = mem.search("what is going on with payments?",
+                  agent="oncall_sre", task_tags=["incident", "latency"])
+for h in hits:
+    print(h["value"], h["promotion"]["lane"])   # every result explains WHY it was served
+```
+
+No profile registered? `search()` is plain hybrid retrieval (keyword + vector + RRF), so you can adopt gradually.
+
+## Plug into a Hermes agent
+
+Three steps on any machine with Hermes installed:
+
+```bash
+# 1. put the engine in Hermes's environment
+pip install engram-lite
+#    (before the PyPI release, install straight from the repo instead:
+#     pip install git+https://github.com/engrammemory-labs/engram-lite.git)
+
+# 2. install the memory plugin
+hermes plugins install engrammemory-labs/engram-lite/hermes-plugin/engram
+
+# 3. activate and configure it
+hermes memory setup     # choose "engram", answer one question — what is this
+                        # agent? (e.g. "DevOps engineer") — everything else is
+                        # derived; leave it empty for plain memory
+```
+
+Then just `hermes chat`. From that point: every turn is auto-captured through the salience gate, the agent boots with a snapshot of what it already knows, and `memory_search` / `memory_write` / `memory_diagnose` are available as in-loop tools. Restart the agent; it remembers. Point several agents at one `db_path` (a wizard field) and the lane model keeps each one's serving scoped.
+
+Building your own harness instead of Hermes? The same provider is a plain Python class:
+
+```python
+from engram.integrations.hermes import EngramMemoryProvider
+
+memory = EngramMemoryProvider(
+    db_path="~/.engram/memory.db",
+    agent="oncall_sre", persona="on-call SRE", domain="sre-devops",
+    scope_tags=["alert", "incident", "latency", "trace", "mitigation"],
+)
+# call initialize / system_prompt_block / prefetch / sync_turn / shutdown
+```
+
+## The numbers behind the claims
+
+### LoCoMo: long-conversation memory, zero LLM in the memory layer
+
+LLM-judge score (J) on LoCoMo categories 1–4 (1,540 questions), under the evaluation protocol standard in the published literature:
+
+| System | Overall J | Single-hop | Multi-hop | Temporal | Open-domain |
+|---|---|---|---|---|---|
+| Full-context baseline (no memory system) | 72.9 | — | — | — | — |
+| **engram-lite** | **68.3** | **74.8** | **55.3** | **68.5** | 49.0 |
+
+- **Zero LLM calls to build the memory.** Ingesting the 5,882-turn corpus costs $0 and takes about 2 minutes, fully local.
+- On the adversarial category (trick questions where the right answer is "no information"): J 64.6. Honest abstention is a feature, not a failure mode.
+- Retrieval alone (no LLM anywhere): 85.7% of questions have their evidence served in the top-30, vs 59.9% for grep over the raw transcript. Two from-scratch runs produce the identical digest.
+- Answerer and judge: claude-haiku-4-5 (measured judge sensitivity: ±2 for prompt wording, ±0 for judge model). LoCoMo: Maharana et al. ([arXiv:2402.17753](https://arxiv.org/abs/2402.17753)). Full protocol notes in [`benchmarks/locomo/`](https://github.com/engrammemory-labs/engram-lite/blob/main/benchmarks/locomo/README.md).
+
+### What memory costs your prompt, per turn
+
+Measured on a real 15-agent shared store (92 memories, 105 probes), same recall quality question for all three strategies:
+
+| Strategy | Tokens injected/turn | Cross-domain leaks | Out-of-lane abstention |
+|---|---|---|---|
+| Whole memory file in the prompt | 2,071 | 85 per turn | 0% |
+| Keyword top-10 over the file | 163 | 5.7 per turn | 0% |
+| **engram-lite conditioned serving** | **29** | **0** | **100%** |
+
+And the memory layer itself bills nothing: zero LLM calls at capture, zero at serving. (LLM-extraction pipelines spend ~2 model calls per captured turn to build memory; engram-lite spends none, and nothing leaves your machine.)
+
+### CAMP-Bench: conditioned serving vs the flat pile
+
+We built a benchmark (CAMP-Bench) with human-authored gold across two engineering domains, 30 cases each: which memory should each role be served, per situation. (An internal benchmark: treat these as our numbers, not independently checkable ones. The LoCoMo results above are the externally replayable ones.) **This library, measured on its real read path** (three-signal retrieval funnel + lane promotion, offline stub embedder, k=3):
+
+| domain | precision | recall | profile discrimination | correct abstention |
+|---|---|---|---|---|
+| sre-devops | 0.63 | 0.87 | 0.91 | 6/6 |
+| backend-engineering | 0.56 | 0.85 | 0.92 | 3/3 |
+| *flat serving (any domain)* | *~0.23* | *~0.55* | *0.00* | *0* |
+
+Downstream, with a live agent and an independent judge on the same data, the lane model took task success from **40% (flat) to 80%** and cut hallucination from **33% to 10%**, while an agent with no memory scored 17%. Flat memory hallucinated more than no memory at all: **wrong memory is worse than no memory.**
+
+## How serving works (the lane model)
+
+Each fact carries tags (extracted at write time from your profiles' vocabulary, or passed explicitly). At recall time:
+
+```
+lane = task_tags ∩ agent.scope_tags          # what this situation is about
+                                             # that THIS agent owns
+empty lane  -> serve nothing (honest abstention)
+otherwise   -> score facts by IDF-weighted overlap with the lane,
+               gate out other domains, apply a relevance floor,
+               serve at most k (fewer when little truly fits)
+```
+
+Deterministic, microseconds, no LLM in the read path, and every served fact carries `why` (the lane tags it hit and its score).
+
+## Development
+
+```bash
+pip install -e ".[dev]"
+ruff check src
+```
+
+## License
+Apache-2.0.

engram_lite-0.1.0/README.md

@@ -0,0 +1,166 @@
+# engram-lite
+
+**Local memory for AI agents, served by persona, domain, and task. Survives restarts. One machine, many agents.**
+
+Most agent memory is a shared pile: every agent gets the same top-k for the same query. In our benchmark (methodology below), an agent fed flat shared memory **hallucinated more than an agent given no memory at all.** Serving each agent the memory that fits **who it is** (its persona, its domain, the task in front of it) **doubled task success (40% to 80%)** and **cut hallucination to a third.**
+
+And the memory layer never calls an LLM. On **LoCoMo** — the long-conversation memory benchmark — engram-lite scores **J 68.3** with **zero LLM calls and $0 to build the memory**. The protocol, footnotes, and the full comparison against the leading LLM-based memory systems live in [`benchmarks/locomo/`](https://github.com/engrammemory-labs/engram-lite/blob/main/benchmarks/locomo/README.md).
+
+engram-lite is that serving layer, small enough to run on your laptop:
+
+- **Conditioned serving.** Register a profile per agent (persona / domain / scope). The same question serves your on-call SRE agent the alert, the trace, and the mitigation runbook, and serves your release agent the deploy, the rollback, and the freeze policy. Out-of-lane questions correctly serve nothing.
+- **Restart-proof.** One SQLite file holds everything. Your agent restarts tomorrow and picks up exactly what it knew.
+- **Self-cleaning.** A salience gate skips junk (code, command output, questions). New facts are de-duplicated, updates supersede old versions, stale facts expire, and the store stays bounded.
+- **Zero infrastructure.** No server, no cloud, no accounts, no telemetry. SQLite plus a local embedding model. `pip install` and you are running.
+- **Explainable forgetting.** Every drop, merge, truncation, and abstention is recorded with the rule that fired (`Memory.decisions()`). "Why don't you remember X?" has an answer — something an LLM-written memory can't give you, because its keep/drop decisions live inside model weights.
+- **Deterministic and replayable.** Same transcript in, same memory out — byte-identical across rebuilds. Memory failures can be diffed, bisected, and regression-gated like any other bug.
+- **Plugs into what you build with.** A Hermes memory provider, or three lines of Python.
+
+> engram-lite is single-machine by design. A team/hosted edition, **engram-core**, is in development — reach out if that is what you need.
+
+## Install
+
+```bash
+pip install engram-lite
+```
+
+Or Docker (memory persists in the named volume):
+
+```bash
+docker build -t engram-lite .
+docker run -it --rm -v engram-data:/data engram-lite demo
+```
+
+## 60 seconds: see conditioned memory work
+
+```bash
+python examples/restart_demo.py    # two agents remember
+python examples/restart_demo.py    # run it AGAIN: memory survived, and the same
+                                   # question serves each agent different memory
+```
+
+## Use it from Python
+
+```python
+from engram import Memory
+
+mem = Memory("~/.engram/memory.db", origin_tool="oncall_sre")
+
+# who is this agent? (this is what makes serving conditioned)
+mem.register_profile(
+    "oncall_sre", persona="on-call SRE", domain="sre-devops",
+    scope_tags=["alert", "incident", "latency", "trace", "mitigation", "db"],
+)
+
+mem.remember("Alert PAGE-99 fired: payments p99 latency at 2.3s")   # gated save
+# conversational input? pass who said it and when — every extracted fragment
+# keeps the anchors, and relative dates resolve to absolute ("last Friday"
+# becomes "[= Friday, 5 May 2023]") with calendar math, no LLM:
+mem.remember("The pool was resized last Friday. We watch it daily now.",
+             speaker="Raj", when="8 May, 2023")
+hits = mem.search("what is going on with payments?",
+                  agent="oncall_sre", task_tags=["incident", "latency"])
+for h in hits:
+    print(h["value"], h["promotion"]["lane"])   # every result explains WHY it was served
+```
+
+No profile registered? `search()` is plain hybrid retrieval (keyword + vector + RRF), so you can adopt gradually.
+
+## Plug into a Hermes agent
+
+Three steps on any machine with Hermes installed:
+
+```bash
+# 1. put the engine in Hermes's environment
+pip install engram-lite
+#    (before the PyPI release, install straight from the repo instead:
+#     pip install git+https://github.com/engrammemory-labs/engram-lite.git)
+
+# 2. install the memory plugin
+hermes plugins install engrammemory-labs/engram-lite/hermes-plugin/engram
+
+# 3. activate and configure it
+hermes memory setup     # choose "engram", answer one question — what is this
+                        # agent? (e.g. "DevOps engineer") — everything else is
+                        # derived; leave it empty for plain memory
+```
+
+Then just `hermes chat`. From that point: every turn is auto-captured through the salience gate, the agent boots with a snapshot of what it already knows, and `memory_search` / `memory_write` / `memory_diagnose` are available as in-loop tools. Restart the agent; it remembers. Point several agents at one `db_path` (a wizard field) and the lane model keeps each one's serving scoped.
+
+Building your own harness instead of Hermes? The same provider is a plain Python class:
+
+```python
+from engram.integrations.hermes import EngramMemoryProvider
+
+memory = EngramMemoryProvider(
+    db_path="~/.engram/memory.db",
+    agent="oncall_sre", persona="on-call SRE", domain="sre-devops",
+    scope_tags=["alert", "incident", "latency", "trace", "mitigation"],
+)
+# call initialize / system_prompt_block / prefetch / sync_turn / shutdown
+```
+
+## The numbers behind the claims
+
+### LoCoMo: long-conversation memory, zero LLM in the memory layer
+
+LLM-judge score (J) on LoCoMo categories 1–4 (1,540 questions), under the evaluation protocol standard in the published literature:
+
+| System | Overall J | Single-hop | Multi-hop | Temporal | Open-domain |
+|---|---|---|---|---|---|
+| Full-context baseline (no memory system) | 72.9 | — | — | — | — |
+| **engram-lite** | **68.3** | **74.8** | **55.3** | **68.5** | 49.0 |
+
+- **Zero LLM calls to build the memory.** Ingesting the 5,882-turn corpus costs $0 and takes about 2 minutes, fully local.
+- On the adversarial category (trick questions where the right answer is "no information"): J 64.6. Honest abstention is a feature, not a failure mode.
+- Retrieval alone (no LLM anywhere): 85.7% of questions have their evidence served in the top-30, vs 59.9% for grep over the raw transcript. Two from-scratch runs produce the identical digest.
+- Answerer and judge: claude-haiku-4-5 (measured judge sensitivity: ±2 for prompt wording, ±0 for judge model). LoCoMo: Maharana et al. ([arXiv:2402.17753](https://arxiv.org/abs/2402.17753)). Full protocol notes in [`benchmarks/locomo/`](https://github.com/engrammemory-labs/engram-lite/blob/main/benchmarks/locomo/README.md).
+
+### What memory costs your prompt, per turn
+
+Measured on a real 15-agent shared store (92 memories, 105 probes), same recall quality question for all three strategies:
+
+| Strategy | Tokens injected/turn | Cross-domain leaks | Out-of-lane abstention |
+|---|---|---|---|
+| Whole memory file in the prompt | 2,071 | 85 per turn | 0% |
+| Keyword top-10 over the file | 163 | 5.7 per turn | 0% |
+| **engram-lite conditioned serving** | **29** | **0** | **100%** |
+
+And the memory layer itself bills nothing: zero LLM calls at capture, zero at serving. (LLM-extraction pipelines spend ~2 model calls per captured turn to build memory; engram-lite spends none, and nothing leaves your machine.)
+
+### CAMP-Bench: conditioned serving vs the flat pile
+
+We built a benchmark (CAMP-Bench) with human-authored gold across two engineering domains, 30 cases each: which memory should each role be served, per situation. (An internal benchmark: treat these as our numbers, not independently checkable ones. The LoCoMo results above are the externally replayable ones.) **This library, measured on its real read path** (three-signal retrieval funnel + lane promotion, offline stub embedder, k=3):
+
+| domain | precision | recall | profile discrimination | correct abstention |
+|---|---|---|---|---|
+| sre-devops | 0.63 | 0.87 | 0.91 | 6/6 |
+| backend-engineering | 0.56 | 0.85 | 0.92 | 3/3 |
+| *flat serving (any domain)* | *~0.23* | *~0.55* | *0.00* | *0* |
+
+Downstream, with a live agent and an independent judge on the same data, the lane model took task success from **40% (flat) to 80%** and cut hallucination from **33% to 10%**, while an agent with no memory scored 17%. Flat memory hallucinated more than no memory at all: **wrong memory is worse than no memory.**
+
+## How serving works (the lane model)
+
+Each fact carries tags (extracted at write time from your profiles' vocabulary, or passed explicitly). At recall time:
+
+```
+lane = task_tags ∩ agent.scope_tags          # what this situation is about
+                                             # that THIS agent owns
+empty lane  -> serve nothing (honest abstention)
+otherwise   -> score facts by IDF-weighted overlap with the lane,
+               gate out other domains, apply a relevance floor,
+               serve at most k (fewer when little truly fits)
+```
+
+Deterministic, microseconds, no LLM in the read path, and every served fact carries `why` (the lane tags it hit and its score).
+
+## Development
+
+```bash
+pip install -e ".[dev]"
+ruff check src
+```
+
+## License
+Apache-2.0.