reverie-memory 0.1.0a1__tar.gz

reverie_memory-0.1.0a1/.gitignore

@@ -0,0 +1,35 @@
+# Python
+__pycache__/
+*.py[cod]
+*.egg-info/
+build/
+dist/
+.venv/
+venv/
+
+# Tooling caches
+.pytest_cache/
+.ipynb_checkpoints/
+.coverage
+htmlcov/
+
+# Local editor / agent settings
+.claude/settings.local.json
+
+# LaTeX build products (paper/ is compiled elsewhere, e.g. Overleaf)
+paper/*.aux
+paper/*.bbl
+paper/*.blg
+paper/*.log
+paper/*.out
+paper/*.fdb_latexmk
+paper/*.fls
+paper/*.synctex.gz
+paper/main.pdf
+
+# OS
+.DS_Store
+Thumbs.db
+
+# Build outputs (PDF, arXiv tarball) -- regenerate, do not commit
+paper/dist/

reverie_memory-0.1.0a1/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.

reverie_memory-0.1.0a1/PKG-INFO

@@ -0,0 +1,211 @@
+Metadata-Version: 2.5
+Name: reverie-memory
+Version: 0.1.0a1
+Summary: Precise precedent retrieval for AI agents, with a human in the loop.
+Project-URL: Homepage, https://github.com/cinostroza/reverie
+Project-URL: Repository, https://github.com/cinostroza/reverie
+Project-URL: Issues, https://github.com/cinostroza/reverie/issues
+Author: Reverie contributors
+License-Expression: Apache-2.0
+License-File: LICENSE
+Keywords: agent,knowledge-graph,llm,mcp,memory
+Classifier: Development Status :: 3 - Alpha
+Classifier: Intended Audience :: Developers
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Programming Language :: Python :: 3.13
+Classifier: Programming Language :: Python :: 3.14
+Requires-Python: >=3.11
+Provides-Extra: dev
+Requires-Dist: pytest-cov>=4.0; extra == 'dev'
+Requires-Dist: pytest>=8.0; extra == 'dev'
+Provides-Extra: experiments
+Requires-Dist: ipykernel>=6.0; extra == 'experiments'
+Requires-Dist: jupyter>=1.0; extra == 'experiments'
+Requires-Dist: matplotlib>=3.8; extra == 'experiments'
+Requires-Dist: networkx>=3.0; extra == 'experiments'
+Requires-Dist: numpy>=1.26; extra == 'experiments'
+Requires-Dist: pandas>=2.0; extra == 'experiments'
+Requires-Dist: scipy>=1.11; extra == 'experiments'
+Description-Content-Type: text/markdown
+
+# Reverie
+
+**Precise precedent retrieval for AI agents, with a human in the loop.**
+
+> ⚠️ **Pre-alpha, and the thesis changed.** This project started out claiming that
+> *outcome attribution* — memories earning a track record and demoting themselves when
+> they hurt — was the differentiator. **We built it, measured it three ways, and it does
+> not work.** See [E7/E9](https://github.com/cinostroza/reverie/blob/main/experiments/README.md). What does work is narrower and better
+> evidenced. Numbers below are from synthetic benchmarks against an honest baseline, not
+> from production agents. Read [Known limitations](#known-limitations) first.
+
+---
+
+Most agent memory optimises **recall fidelity**: can the system retrieve the thing you
+told it. Reverie optimises **precedent precision**: given what this agent is doing right
+now, surface the specific past case that applies — and let a human correct the record
+when it is wrong.
+
+Two claims, both measured on `reverie-bench` over a 72-context on-call landscape at 600
+tasks, against budget-matched history-stuffing (what people actually do today):
+
+| | Reverie | Naive replay | Reverie + human review |
+|---|---|---|---|
+| success rate | 0.459 ±0.020 | 0.402 ±0.022 | **0.640 ±0.019** |
+| tokens per prompt | 775 | 1,146 | **708** |
+
+Paired over 16 identical seeds, so seed variance cancels:
+**Reverie − replay = +0.056 ±0.032 (significant)**;
+**human review − Reverie = +0.181 ±0.026 (significant)**.
+
+The learning curves are the real story: Reverie **rises** with accumulated history
+(0.300 → 0.439 from 200 to 600 tasks in the committed E6 run) while a recency window is
+**flat** (0.402 → 0.381), because a window cannot reach past itself however much history
+piles up behind it.
+
+**Everything except recall happens out of band.** Consolidation, linking, and human
+review run between sessions. Nothing Reverie does makes the current task slower; it
+exists to make the next one better.
+
+## Install
+
+```bash
+pip install reverie-memory            # core: zero dependencies, stdlib sqlite3 only
+```
+
+The distribution is `reverie-memory`; the import and the CLI are both `reverie`.
+To reproduce the experiments, work from a clone:
+
+```bash
+pip install -e ".[experiments,dev]"   # + numpy/scipy/pandas/matplotlib, pytest
+```
+
+## Use
+
+```python
+from reverie import Reverie
+
+mem = Reverie(scope="agent:deploy-bot")        # defaults to ~/.reverie/memory.db
+
+result = mem.recall("migrate orders table, add shipping_status")
+print(result.brief)                            # budgeted, templated, no LLM call
+
+with mem.episode("migrate orders", recall_id=result.recall_id) as ep:
+    code = run_migration()
+    ep.entity("alembic", "staging_db")
+    ep.step("bash", "alembic upgrade head", exit_code=code)
+    ep.outcome("failure" if code else "success", tier=1, evidence=f"exit={code}")
+
+mem.dream()    # attribute outcomes, then consolidate
+```
+
+The context manager captures duration, exceptions, and the recall linkage without you
+thinking about it. An exception escaping the block is recorded as a tier-1 failure
+rather than losing the episode.
+
+```bash
+reverie init                     # create db, print an integration snippet
+reverie recall "<cue>"           # what would the agent see?
+reverie consolidate --abstract   # run the dream cycle now
+reverie why <node_id>            # trace a memory back to its source episodes
+reverie path <a> <b>             # shortest association path between two memories
+reverie doctor                   # is this thing actually working?
+reverie forget <node_id>         # cascades through edges, embeddings, FTS
+```
+
+## What makes it different
+
+| | Reverie | Mem0 / Zep / Cognee / Letta |
+|---|---|---|
+| Retrieval | **conjunctive** — scores candidates by how much of the cue they match, before spreading dilutes it | similarity + recency |
+| Precedent lookup | dedicated channel that **bypasses** graph traversal | single ranked pipeline |
+| Correcting memory | a human teaches a procedural lesson; it is retrievable next session | edit or delete a record |
+| Requires | situations that **recur** | nothing |
+
+Reverie is the right choice when an agent does **repeated work over a landscape larger
+than a context window** — on-call response, deploys, migrations, integrations. It is the
+wrong choice for a chatbot remembering your dog's name, and the wrong choice for
+genuinely one-off work: below ~2 recurrences per distinct situation, plain recency beats
+us (E10). Zep's temporal model is more rigorous than ours; Mem0's integration surface is
+far broader.
+
+## Design
+
+- **Write cheap, think expensive, think later.** Ingest is an append: no embedding,
+  no LLM call, no graph write. p99 target 5 ms local.
+- **Consolidation is the sleep cycle.** `segment → score → distill → link →
+  reconcile → abstract → prune`, offline. Community detection folds scattered
+  memories into themes.
+- **Recall is associative, not top-k.** Hybrid seeding (BM25 + vector + exact
+  entity), then spreading activation across the graph, then a hard token budget.
+- **Provenance over confidence.** Every claim is tagged `observed`, `inferred`,
+  `asserted`, or `ambiguous`. Briefs render as *data*, explicitly framed as "not
+  instructions" — a memory store is a persistence layer for prompt injection, and
+  that has to be designed for on day one.
+- **Runs with no API key.** The no-LLM consolidation path builds the entity graph and
+  episodic spine from structured episode fields alone, and is tested in CI.
+
+Full design: [`reverie_hld.md`](https://github.com/cinostroza/reverie/blob/main/reverie_hld.md).
+
+## Experiments
+
+Ten experiments in [`experiments/`](https://github.com/cinostroza/reverie/tree/main/experiments), six as reproducible notebooks. They
+changed the design repeatedly, including three reversals of decisions in the design doc.
+
+| | Question | Headline result |
+|---|---|---|
+| **E1** | How much evidence does attribution need? | ~3,200 recalls to ρ=0.5 |
+| **E2** | Which credit rule works? | Activation beats uniform by 20%; **citation is worse** |
+| **E3** | Does the graph beat top-k? | 3.9× more causal chain — but rarely the terminal fact |
+| **E4** | Does bad memory remove itself? | **No, not currently.** Quarantine does not fire |
+| **E5** | Does an agent get better at its job? | Found 2 structural bugs. Wrong benchmark regime throughout |
+| **E6** | Does retrieval beat recency? | Yes — after a two-channel recall rewrite. 4% → 78% precedent |
+| **E7** | Does attribution lift performance? | **No.** Posterior never leaves the prior |
+| **E8** | Does human review help? | Not as first built — a retrieval bug, not a bad idea |
+| **E9** | …after fixing the plumbing? | **+0.227.** Attribution still zero |
+| **E10** | Where does it break? | Repetition density is the governing variable |
+
+## Known limitations
+
+Stated up front, because they are the first things a careful reader will look for.
+
+1. **Outcome attribution does not work.** Zero lift in three independent tests, and the
+   constraint is arithmetic rather than a tuning failure: total credit mass equals the
+   number of outcomes, so a graph with more memories than outcomes cannot give any one
+   memory enough evidence to rank on. The code ships, disabled-by-default is under
+   consideration, and it is documented as a negative result rather than removed.
+2. **The win is conditional on repetition.** At 8.3 tasks per distinct context we beat
+   recency; at 2.0 we lose; at 1.0 we are at chance. Measure your repetition density
+   before adopting this.
+3. **An over-general human lesson is worse than no human at all.** A reviewer who drops
+   a feature that mattered drives performance *below* the no-review baseline
+   (0.411 vs 0.456). Guarding against this is the top unbuilt item. Reviewers who are
+   simply *wrong* — even adversarially, even 50% of the time — are handled fine.
+4. **No LLM distiller.** `NullDistiller` produces no procedural or semantic memories, so
+   everything measured rests on episodic recall plus human lessons.
+5. **Recall does not reliably surface 3-hop facts** (E3). The associative demo in the
+   design doc overclaims relative to what default `recall` does today.
+6. **Quarantine does not work.** The E4 scenario that once removed a poisoned memory in
+   23 recalls now never fires, and held-out seeds put detection at 5/20 even when tuned.
+   Design goal G8 — "memory can be shown to be wrong and removed automatically" — is not
+   currently satisfied.
+7. **Single-tenant.** No org/RBAC layer. The SQLite file is sensitive — it holds
+   redacted tool output, and redaction is best-effort.
+
+## Development
+
+```bash
+pytest -q                              # 76 tests
+python experiments/_build_notebooks.py # regenerate E1-E4
+python experiments/_build_e5.py        # E5
+python experiments/_build_e6.py        # E6
+```
+
+The core has **no runtime dependencies**. numpy/scipy are experiment-side only, which
+is why the Beta quantile in `reverie/_math.py` is implemented from scratch and
+verified against `scipy.stats.beta` to 1e-9 in the test suite.
+
+## License
+
+Apache-2.0.

reverie_memory-0.1.0a1/README.md

@@ -0,0 +1,180 @@
+# Reverie
+
+**Precise precedent retrieval for AI agents, with a human in the loop.**
+
+> ⚠️ **Pre-alpha, and the thesis changed.** This project started out claiming that
+> *outcome attribution* — memories earning a track record and demoting themselves when
+> they hurt — was the differentiator. **We built it, measured it three ways, and it does
+> not work.** See [E7/E9](https://github.com/cinostroza/reverie/blob/main/experiments/README.md). What does work is narrower and better
+> evidenced. Numbers below are from synthetic benchmarks against an honest baseline, not
+> from production agents. Read [Known limitations](#known-limitations) first.
+
+---
+
+Most agent memory optimises **recall fidelity**: can the system retrieve the thing you
+told it. Reverie optimises **precedent precision**: given what this agent is doing right
+now, surface the specific past case that applies — and let a human correct the record
+when it is wrong.
+
+Two claims, both measured on `reverie-bench` over a 72-context on-call landscape at 600
+tasks, against budget-matched history-stuffing (what people actually do today):
+
+| | Reverie | Naive replay | Reverie + human review |
+|---|---|---|---|
+| success rate | 0.459 ±0.020 | 0.402 ±0.022 | **0.640 ±0.019** |
+| tokens per prompt | 775 | 1,146 | **708** |
+
+Paired over 16 identical seeds, so seed variance cancels:
+**Reverie − replay = +0.056 ±0.032 (significant)**;
+**human review − Reverie = +0.181 ±0.026 (significant)**.
+
+The learning curves are the real story: Reverie **rises** with accumulated history
+(0.300 → 0.439 from 200 to 600 tasks in the committed E6 run) while a recency window is
+**flat** (0.402 → 0.381), because a window cannot reach past itself however much history
+piles up behind it.
+
+**Everything except recall happens out of band.** Consolidation, linking, and human
+review run between sessions. Nothing Reverie does makes the current task slower; it
+exists to make the next one better.
+
+## Install
+
+```bash
+pip install reverie-memory            # core: zero dependencies, stdlib sqlite3 only
+```
+
+The distribution is `reverie-memory`; the import and the CLI are both `reverie`.
+To reproduce the experiments, work from a clone:
+
+```bash
+pip install -e ".[experiments,dev]"   # + numpy/scipy/pandas/matplotlib, pytest
+```
+
+## Use
+
+```python
+from reverie import Reverie
+
+mem = Reverie(scope="agent:deploy-bot")        # defaults to ~/.reverie/memory.db
+
+result = mem.recall("migrate orders table, add shipping_status")
+print(result.brief)                            # budgeted, templated, no LLM call
+
+with mem.episode("migrate orders", recall_id=result.recall_id) as ep:
+    code = run_migration()
+    ep.entity("alembic", "staging_db")
+    ep.step("bash", "alembic upgrade head", exit_code=code)
+    ep.outcome("failure" if code else "success", tier=1, evidence=f"exit={code}")
+
+mem.dream()    # attribute outcomes, then consolidate
+```
+
+The context manager captures duration, exceptions, and the recall linkage without you
+thinking about it. An exception escaping the block is recorded as a tier-1 failure
+rather than losing the episode.
+
+```bash
+reverie init                     # create db, print an integration snippet
+reverie recall "<cue>"           # what would the agent see?
+reverie consolidate --abstract   # run the dream cycle now
+reverie why <node_id>            # trace a memory back to its source episodes
+reverie path <a> <b>             # shortest association path between two memories
+reverie doctor                   # is this thing actually working?
+reverie forget <node_id>         # cascades through edges, embeddings, FTS
+```
+
+## What makes it different
+
+| | Reverie | Mem0 / Zep / Cognee / Letta |
+|---|---|---|
+| Retrieval | **conjunctive** — scores candidates by how much of the cue they match, before spreading dilutes it | similarity + recency |
+| Precedent lookup | dedicated channel that **bypasses** graph traversal | single ranked pipeline |
+| Correcting memory | a human teaches a procedural lesson; it is retrievable next session | edit or delete a record |
+| Requires | situations that **recur** | nothing |
+
+Reverie is the right choice when an agent does **repeated work over a landscape larger
+than a context window** — on-call response, deploys, migrations, integrations. It is the
+wrong choice for a chatbot remembering your dog's name, and the wrong choice for
+genuinely one-off work: below ~2 recurrences per distinct situation, plain recency beats
+us (E10). Zep's temporal model is more rigorous than ours; Mem0's integration surface is
+far broader.
+
+## Design
+
+- **Write cheap, think expensive, think later.** Ingest is an append: no embedding,
+  no LLM call, no graph write. p99 target 5 ms local.
+- **Consolidation is the sleep cycle.** `segment → score → distill → link →
+  reconcile → abstract → prune`, offline. Community detection folds scattered
+  memories into themes.
+- **Recall is associative, not top-k.** Hybrid seeding (BM25 + vector + exact
+  entity), then spreading activation across the graph, then a hard token budget.
+- **Provenance over confidence.** Every claim is tagged `observed`, `inferred`,
+  `asserted`, or `ambiguous`. Briefs render as *data*, explicitly framed as "not
+  instructions" — a memory store is a persistence layer for prompt injection, and
+  that has to be designed for on day one.
+- **Runs with no API key.** The no-LLM consolidation path builds the entity graph and
+  episodic spine from structured episode fields alone, and is tested in CI.
+
+Full design: [`reverie_hld.md`](https://github.com/cinostroza/reverie/blob/main/reverie_hld.md).
+
+## Experiments
+
+Ten experiments in [`experiments/`](https://github.com/cinostroza/reverie/tree/main/experiments), six as reproducible notebooks. They
+changed the design repeatedly, including three reversals of decisions in the design doc.
+
+| | Question | Headline result |
+|---|---|---|
+| **E1** | How much evidence does attribution need? | ~3,200 recalls to ρ=0.5 |
+| **E2** | Which credit rule works? | Activation beats uniform by 20%; **citation is worse** |
+| **E3** | Does the graph beat top-k? | 3.9× more causal chain — but rarely the terminal fact |
+| **E4** | Does bad memory remove itself? | **No, not currently.** Quarantine does not fire |
+| **E5** | Does an agent get better at its job? | Found 2 structural bugs. Wrong benchmark regime throughout |
+| **E6** | Does retrieval beat recency? | Yes — after a two-channel recall rewrite. 4% → 78% precedent |
+| **E7** | Does attribution lift performance? | **No.** Posterior never leaves the prior |
+| **E8** | Does human review help? | Not as first built — a retrieval bug, not a bad idea |
+| **E9** | …after fixing the plumbing? | **+0.227.** Attribution still zero |
+| **E10** | Where does it break? | Repetition density is the governing variable |
+
+## Known limitations
+
+Stated up front, because they are the first things a careful reader will look for.
+
+1. **Outcome attribution does not work.** Zero lift in three independent tests, and the
+   constraint is arithmetic rather than a tuning failure: total credit mass equals the
+   number of outcomes, so a graph with more memories than outcomes cannot give any one
+   memory enough evidence to rank on. The code ships, disabled-by-default is under
+   consideration, and it is documented as a negative result rather than removed.
+2. **The win is conditional on repetition.** At 8.3 tasks per distinct context we beat
+   recency; at 2.0 we lose; at 1.0 we are at chance. Measure your repetition density
+   before adopting this.
+3. **An over-general human lesson is worse than no human at all.** A reviewer who drops
+   a feature that mattered drives performance *below* the no-review baseline
+   (0.411 vs 0.456). Guarding against this is the top unbuilt item. Reviewers who are
+   simply *wrong* — even adversarially, even 50% of the time — are handled fine.
+4. **No LLM distiller.** `NullDistiller` produces no procedural or semantic memories, so
+   everything measured rests on episodic recall plus human lessons.
+5. **Recall does not reliably surface 3-hop facts** (E3). The associative demo in the
+   design doc overclaims relative to what default `recall` does today.
+6. **Quarantine does not work.** The E4 scenario that once removed a poisoned memory in
+   23 recalls now never fires, and held-out seeds put detection at 5/20 even when tuned.
+   Design goal G8 — "memory can be shown to be wrong and removed automatically" — is not
+   currently satisfied.
+7. **Single-tenant.** No org/RBAC layer. The SQLite file is sensitive — it holds
+   redacted tool output, and redaction is best-effort.
+
+## Development
+
+```bash
+pytest -q                              # 76 tests
+python experiments/_build_notebooks.py # regenerate E1-E4
+python experiments/_build_e5.py        # E5
+python experiments/_build_e6.py        # E6
+```
+
+The core has **no runtime dependencies**. numpy/scipy are experiment-side only, which
+is why the Beta quantile in `reverie/_math.py` is implemented from scratch and
+verified against `scipy.stats.beta` to 1e-9 in the test suite.
+
+## License
+
+Apache-2.0.