3tears-agent-memory 0.14.0__tar.gz

3tears_agent_memory-0.14.0/.gitignore

@@ -0,0 +1,216 @@
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[codz]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py.cover
+.hypothesis/
+.pytest_cache/
+cover/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+.pybuilder/
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+#   For a library or package, you might want to ignore these files since the code is
+#   intended to run in multiple environments; otherwise, check them in:
+# .python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# UV
+#   Similar to Pipfile.lock, it is generally recommended to include uv.lock in version control.
+#   This is especially recommended for binary packages to ensure reproducibility, and is more
+#   commonly ignored for libraries.
+#uv.lock
+
+# poetry
+#   Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
+#   This is especially recommended for binary packages to ensure reproducibility, and is more
+#   commonly ignored for libraries.
+#   https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
+#poetry.lock
+#poetry.toml
+
+# pdm
+#   Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
+#   pdm recommends including project-wide configuration in pdm.toml, but excluding .pdm-python.
+#   https://pdm-project.org/en/latest/usage/project/#working-with-version-control
+#pdm.lock
+#pdm.toml
+.pdm-python
+.pdm-build/
+
+# pixi
+#   Similar to Pipfile.lock, it is generally recommended to include pixi.lock in version control.
+#pixi.lock
+#   Pixi creates a virtual environment in the .pixi directory, just like venv module creates one
+#   in the .venv directory. It is recommended not to include this directory in version control.
+.pixi
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.envrc
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+# pytype static type analyzer
+.pytype/
+
+# Cython debug symbols
+cython_debug/
+
+# PyCharm
+#  JetBrains specific template is maintained in a separate JetBrains.gitignore that can
+#  be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
+#  and can be added to the global gitignore or merged into this file.  For a more nuclear
+#  option (not recommended) you can uncomment the following to ignore the entire idea folder.
+#.idea/
+
+# Abstra
+# Abstra is an AI-powered process automation framework.
+# Ignore directories containing user credentials, local state, and settings.
+# Learn more at https://abstra.io/docs
+.abstra/
+
+# Visual Studio Code
+#  Visual Studio Code specific template is maintained in a separate VisualStudioCode.gitignore 
+#  that can be found at https://github.com/github/gitignore/blob/main/Global/VisualStudioCode.gitignore
+#  and can be added to the global gitignore or merged into this file. However, if you prefer, 
+#  you could uncomment the following to ignore the entire vscode folder
+# .vscode/
+
+# Ruff stuff:
+.ruff_cache/
+
+# PyPI configuration file
+.pypirc
+
+# Cursor
+#  Cursor is an AI-powered code editor. `.cursorignore` specifies files/directories to
+#  exclude from AI features like autocomplete and code analysis. Recommended for sensitive data
+#  refer to https://docs.cursor.com/context/ignore-files
+.cursorignore
+.cursorindexingignore
+
+# Marimo
+marimo/_static/
+marimo/_lsp/
+__marimo__/
+
+# Claude Code local state
+.claude/
+
+# prawduct session evidence (local governance artifacts, never shipped)
+.prawduct/
+
+# macOS folder metadata
+.DS_Store

3tears_agent_memory-0.14.0/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Mark Pace
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

3tears_agent_memory-0.14.0/PKG-INFO

@@ -0,0 +1,231 @@
+Metadata-Version: 2.4
+Name: 3tears-agent-memory
+Version: 0.14.0
+Summary: Memory system for LLM agents -- extraction, retrieval, hybrid search, and MMR reranking
+Project-URL: Repository, https://github.com/pacepace/3tears
+Author: pace
+License-Expression: MIT
+License-File: LICENSE
+Classifier: Development Status :: 3 - Alpha
+Classifier: Framework :: AsyncIO
+Classifier: Intended Audience :: Developers
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.14
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Classifier: Topic :: Software Development :: Libraries
+Classifier: Typing :: Typed
+Requires-Python: >=3.14
+Requires-Dist: 3tears
+Requires-Dist: 3tears-agent-acl
+Requires-Dist: 3tears-langgraph
+Requires-Dist: 3tears-observe
+Requires-Dist: langchain-core
+Requires-Dist: pgvector
+Description-Content-Type: text/markdown
+
+# 3tears Agent Memory
+
+Memory system for LLM agents. Handles extraction of memorable facts from conversations, hybrid retrieval (semantic + full-text + recency), and memory lifecycle management.
+
+Part of the [3tears](https://github.com/pacepace/3tears) framework.
+
+## Installation
+
+```bash
+pip install 3tears-agent-memory
+```
+
+## Components
+
+### Collections are the single entry point for memory-table SQL
+
+Every memory-table write, single-row read, batch read, and hybrid-search query goes through one of four `BaseCollection` subclasses; no consumer of this package holds an `asyncpg.Pool` reference directly.
+
+- `MemoriesCollection` -- `memories` table. CRUD through `get` / `save_entity` / `delete`; complex queries through `hybrid_search`, `search_by_ids`, `search_by_semantic`, `search_by_fts`, `find_similar_for_dedup`, `count_by_user`, `fetch_content_for_recall`.
+- `MediaCollection` -- `media` parent table. CRUD only.
+- `MediaContentCollection` -- `media_content` child table. CRUD + `hybrid_search`, `search_by_ids`, `search_by_semantic`, `search_by_fts`, `fetch_content_for_recall`.
+- `MemoryChunkCollection` -- `memory_chunks` child table. CRUD + `hybrid_search`, `search_by_ids`, `search_by_semantic`, `fetch_content_for_recall`.
+
+All four resolve their L3 pool through `CollectionRegistry` (same pattern `ConversationCollection` uses); an L1 `SQLiteBackend` attached to the registry populates on `save_entity` and serves subsequent by-id `get` calls without an L3 round-trip. The `media` / `media_content` / `memory_chunks` tables are introduced by migrations v006 / v007.
+
+Hybrid-search methods carry documented `# cache-bypass: <reason>` inline comments because the query shape (vector distance, FTS rank, multi-table joins) is not primary-key-addressable and the L1 row cache cannot serve the lookup. Keeping the SQL on the Collection preserves the single entry point. The cache-primitive enforcement walker recognises in-Collection bypass sites as legitimate and reports any bypass that leaks back into `retrieval.py` / `extraction.py` / `tools.py` as a violation.
+
+### MemoryExtractor
+
+Extracts memorable facts from conversation turns. Uses a multi-stage pipeline: candidate extraction via LLM, deduplication against existing memories via embedding similarity, and action resolution (ADD / UPDATE / DELETE).
+
+```python
+from threetears.agent.memory import (
+    MemoriesCollection,
+    MemoryConfig,
+    MemoryExtractor,
+)
+
+extractor = MemoryExtractor(
+    config=MemoryConfig(),
+    embedding_provider=my_embedding_provider,
+    chat_model_factory=my_chat_model_factory,
+    authorizer=authorizer_bundle,
+    memories_collection=memories_collection,
+    summary_callback=on_new_memory,
+)
+
+await extractor.extract(
+    user_id=user_id,
+    conversation_id=conv_id,
+    message_id_source=msg_id,
+    user_message="I just moved to Portland",
+    assistant_response="That's exciting! Portland has great food...",
+    turn_count=5,
+    agent_id=agent_id,
+    customer_id=customer_id,
+)
+```
+
+### MemoryRetriever
+
+Retrieves relevant memories using hybrid search: pgvector semantic similarity, PostgreSQL full-text search, recency decay, and MMR reranking for diversity. Takes the three search-bearing Collections at construction; no pool.
+
+```python
+from threetears.agent.memory import MemoryRetriever, MemoryConfig
+
+retriever = MemoryRetriever(
+    config=MemoryConfig(),
+    embedding_provider=my_embedding_provider,
+    authorizer=authorizer_bundle,
+    memories_collection=memories_collection,
+    media_content_collection=media_content_collection,
+    memory_chunk_collection=memory_chunk_collection,
+)
+
+result = await retriever.retrieve_with_candidates(
+    user_id,
+    "Tell me about Portland",
+    agent_id=agent_id,
+    customer_id=customer_id,
+    caller_user_id=user_id,
+    caller_agent_id=agent_id,
+)
+
+# result.context     -- formatted string for injection into system prompt
+# result.memories    -- raw memory dicts with similarity scores
+# result.media_content -- matched media content
+# result.memory_chunks -- matched document chunks
+```
+
+### Protocols
+
+Implement these to integrate with your infrastructure:
+
+```python
+from threetears.agent.memory import EmbeddingProvider, ChatModelFactory
+
+class MyEmbeddingProvider(EmbeddingProvider):
+    async def embed(self, text: str) -> tuple[list[float], int, UUID]:
+        # Returns (embedding_vector, token_count, model_id)
+        ...
+
+class MyChatModelFactory(ChatModelFactory):
+    async def create_chat_model(self, purpose: str = "extraction"):
+        # Returns a langchain BaseChatModel
+        ...
+```
+
+### Tools
+
+LangChain tools for agent use: memory search, recall, and explicit add. Factories take Collection references; no pool:
+
+```python
+from threetears.agent.memory import (
+    load_memory_add_tool,
+    load_memory_search_tool,
+    load_memory_recall_tool,
+)
+
+search_tool = await load_memory_search_tool(
+    user_id=user_id,
+    embedding_provider=embedding_provider,
+    agent_id=agent_id,
+    customer_id=customer_id,
+    authorizer=authorizer_bundle,
+    memories_collection=memories_collection,
+    media_content_collection=media_content_collection,
+    memory_chunk_collection=memory_chunk_collection,
+)
+recall_tool = await load_memory_recall_tool(
+    user_id=user_id,
+    agent_id=agent_id,
+    customer_id=customer_id,
+    authorizer=authorizer_bundle,
+    memories_collection=memories_collection,
+    media_content_collection=media_content_collection,
+    memory_chunk_collection=memory_chunk_collection,
+)
+add_tool = await load_memory_add_tool(
+    user_id=user_id,
+    conversation_id=conv_id,
+    message_id=msg_id,
+    embedding_provider=embedding_provider,
+    agent_id=agent_id,
+    customer_id=customer_id,
+    authorizer=authorizer_bundle,
+    memories_collection=memories_collection,
+)
+```
+
+### Configuration
+
+```python
+from threetears.agent.memory import MemoryConfig
+
+config = MemoryConfig(
+    similarity_threshold=0.4,      # minimum cosine similarity for retrieval
+    detail_threshold=0.85,         # threshold for including full memory detail
+    context_budget=15,             # max memories in context
+    dedup_threshold=0.85,          # similarity threshold for deduplication
+    max_candidates=10,             # max candidates per extraction
+)
+```
+
+## Database Schema
+
+Requires PostgreSQL with the `pgvector` extension. The package's own migration runner (`threetears.agent.memory.migrations.register`) produces the full schema per agent schema. Registered versions:
+
+- **v001** -- `memories` (PK `memory_id`, pgvector `embedding`, scoping ids, content, summary, lifecycle timestamps).
+- **v002** -- `conversation_memory_refs` (ledger of per-conversation surfaced items).
+- **v003** -- column reconciliation: renames PK and discriminator to match the package code (`id` to `memory_id`, `memory_type` to `type_memory`), drops columns the code does not read (`embedding_model`, `importance`, `metadata`, `date_accessed`), loosens `agent_id`/`customer_id` to NULL.
+- **v004** -- lifecycle + conversation-link columns on `memories` (`conversation_id`, `message_id_source`, `is_deleted`, `media_id`, `date_deleted`, `summary`) with indexes.
+- **v005** -- FTS: `search_vector TSVECTOR` + GIN index + maintenance trigger on `memories`.
+- **v006** -- `media` (parent) + `media_content` (chunked extracted text with embedding + FTS).
+- **v007** -- `memory_chunks` (document-style chunks with heading / page metadata + embedding + FTS).
+
+Every FTS column is trigger-maintained from `content` + `summary` (weighted A/B); callers do not have to populate `search_vector` manually. Integration tests under `tests/integration/` exercise the full chain + every public API surface against `pgvector/pgvector:pg16` via testcontainers.
+
+## RBAC Enforcement
+
+Memory reads, writes, and extractions flow through the unified rbac evaluator in `threetears.agent.acl`. Every (agent, customer) pair is a `memory`-type namespace in the `namespaces` table; each access resolves the namespace and evaluates one of three canonical actions against the caller's `(user_id, agent_id)` pair:
+
+- `memory.read` -- retrieval / search / recall. Guarded on `MemoryRetriever.retrieve*`, `MemoriesCollection.find_by_user`, `MemoriesCollection.find_by_scope`, the `memory_search` + `memory_recall` LangChain tools.
+- `memory.write` -- user-initiated writes. Guarded on `MemoriesCollection.save_memory` and the `memory_add` LangChain tool.
+- `memory.extract` -- agent-internal extraction path. Guarded on `MemoryExtractor.extract`; the owner short-circuit keeps the common case (agent emitting memories on its own namespace) grant-free.
+
+Owner short-circuit: the evaluator allows any action when the calling agent owns the memory namespace. Agent-internal retrieval and extraction therefore work without explicit grants; user-initiated reads and writes require evaluator assignments.
+
+Auto-assignment on first user-write: `memory_add` ensures a `MemoryOwner` assignment for the calling user on their first write (idempotent-by-state; the ensurer only fires when the user has zero memory rows in the target schema). Subsequent writes authorize against the materialized grant; admin-revoked grants stay revoked (the ensurer does not resurrect them).
+
+Wiring shape: every consumer of the memory surface REQUIRES a `MemoryAuthorizerDependencies` bundle exposing:
+
+- `acl_cache` -- shared `threetears.agent.acl.AclCache` instance;
+- `membership_loader` + `grant_loader` -- the evaluator's loaders (`threetears.agent.acl.MembershipLoader` / `GrantLoader`);
+- `namespace_collection` -- three-tier `NamespaceCollection` used to resolve the memory namespace via `get_by_owner_and_customer(namespace_type="memory", owner_agent_id, customer_id)` (create-if-absent flows through `save_entity`);
+- `group_collection` + `group_member_collection` + `role_collection` + `role_assignment_collection` -- the rbac Collections the first-write owner-assignment path uses via `ensure_memory_owner_assignment(...)`.
+
+There is no bypass. Every `MemoriesCollection`, `MemoryRetriever`, `MemoryExtractor`, and LangChain tool factory (`load_memory_search_tool`, `load_memory_add_tool`, `load_memory_recall_tool`) takes the bundle as a required constructor/factory argument; every code path that touches a memory row runs `authorize_memory_access` first. Callers that omit the bundle fail at the type checker and the Python signature boundary.
+
+- Production wiring builds the bundle directly from the agent-side three-tier stack's Collections (`NatsProxyL3Backend`-backed `NamespaceCollection` / `GroupCollection` / ...).
+- Test wiring injects a permissive fixture `permissive_memory_authorizer` (see `tests/conftest.py`) that carries in-memory Collection stand-ins and a permissive evaluator. Fixture usage is explicit in every test file that constructs a memory surface.
+- Back-office / admin tooling that genuinely needs to read or write memories without an identity must construct its own bundle with Collections bound directly to an asyncpg pool; there is no global escape hatch.
+
+See `threetears.agent.memory.authorize` for the full public surface.
+
+The three platform roles (`MemoryOwner` / `MemoryReader` / `MemoryWriter`) carry the canonical action vocabulary. Platform-side migrations seed these roles and backfill the rbac rows required for evaluator resolution.

3tears_agent_memory-0.14.0/README.md

@@ -0,0 +1,206 @@
+# 3tears Agent Memory
+
+Memory system for LLM agents. Handles extraction of memorable facts from conversations, hybrid retrieval (semantic + full-text + recency), and memory lifecycle management.
+
+Part of the [3tears](https://github.com/pacepace/3tears) framework.
+
+## Installation
+
+```bash
+pip install 3tears-agent-memory
+```
+
+## Components
+
+### Collections are the single entry point for memory-table SQL
+
+Every memory-table write, single-row read, batch read, and hybrid-search query goes through one of four `BaseCollection` subclasses; no consumer of this package holds an `asyncpg.Pool` reference directly.
+
+- `MemoriesCollection` -- `memories` table. CRUD through `get` / `save_entity` / `delete`; complex queries through `hybrid_search`, `search_by_ids`, `search_by_semantic`, `search_by_fts`, `find_similar_for_dedup`, `count_by_user`, `fetch_content_for_recall`.
+- `MediaCollection` -- `media` parent table. CRUD only.
+- `MediaContentCollection` -- `media_content` child table. CRUD + `hybrid_search`, `search_by_ids`, `search_by_semantic`, `search_by_fts`, `fetch_content_for_recall`.
+- `MemoryChunkCollection` -- `memory_chunks` child table. CRUD + `hybrid_search`, `search_by_ids`, `search_by_semantic`, `fetch_content_for_recall`.
+
+All four resolve their L3 pool through `CollectionRegistry` (same pattern `ConversationCollection` uses); an L1 `SQLiteBackend` attached to the registry populates on `save_entity` and serves subsequent by-id `get` calls without an L3 round-trip. The `media` / `media_content` / `memory_chunks` tables are introduced by migrations v006 / v007.
+
+Hybrid-search methods carry documented `# cache-bypass: <reason>` inline comments because the query shape (vector distance, FTS rank, multi-table joins) is not primary-key-addressable and the L1 row cache cannot serve the lookup. Keeping the SQL on the Collection preserves the single entry point. The cache-primitive enforcement walker recognises in-Collection bypass sites as legitimate and reports any bypass that leaks back into `retrieval.py` / `extraction.py` / `tools.py` as a violation.
+
+### MemoryExtractor
+
+Extracts memorable facts from conversation turns. Uses a multi-stage pipeline: candidate extraction via LLM, deduplication against existing memories via embedding similarity, and action resolution (ADD / UPDATE / DELETE).
+
+```python
+from threetears.agent.memory import (
+    MemoriesCollection,
+    MemoryConfig,
+    MemoryExtractor,
+)
+
+extractor = MemoryExtractor(
+    config=MemoryConfig(),
+    embedding_provider=my_embedding_provider,
+    chat_model_factory=my_chat_model_factory,
+    authorizer=authorizer_bundle,
+    memories_collection=memories_collection,
+    summary_callback=on_new_memory,
+)
+
+await extractor.extract(
+    user_id=user_id,
+    conversation_id=conv_id,
+    message_id_source=msg_id,
+    user_message="I just moved to Portland",
+    assistant_response="That's exciting! Portland has great food...",
+    turn_count=5,
+    agent_id=agent_id,
+    customer_id=customer_id,
+)
+```
+
+### MemoryRetriever
+
+Retrieves relevant memories using hybrid search: pgvector semantic similarity, PostgreSQL full-text search, recency decay, and MMR reranking for diversity. Takes the three search-bearing Collections at construction; no pool.
+
+```python
+from threetears.agent.memory import MemoryRetriever, MemoryConfig
+
+retriever = MemoryRetriever(
+    config=MemoryConfig(),
+    embedding_provider=my_embedding_provider,
+    authorizer=authorizer_bundle,
+    memories_collection=memories_collection,
+    media_content_collection=media_content_collection,
+    memory_chunk_collection=memory_chunk_collection,
+)
+
+result = await retriever.retrieve_with_candidates(
+    user_id,
+    "Tell me about Portland",
+    agent_id=agent_id,
+    customer_id=customer_id,
+    caller_user_id=user_id,
+    caller_agent_id=agent_id,
+)
+
+# result.context     -- formatted string for injection into system prompt
+# result.memories    -- raw memory dicts with similarity scores
+# result.media_content -- matched media content
+# result.memory_chunks -- matched document chunks
+```
+
+### Protocols
+
+Implement these to integrate with your infrastructure:
+
+```python
+from threetears.agent.memory import EmbeddingProvider, ChatModelFactory
+
+class MyEmbeddingProvider(EmbeddingProvider):
+    async def embed(self, text: str) -> tuple[list[float], int, UUID]:
+        # Returns (embedding_vector, token_count, model_id)
+        ...
+
+class MyChatModelFactory(ChatModelFactory):
+    async def create_chat_model(self, purpose: str = "extraction"):
+        # Returns a langchain BaseChatModel
+        ...
+```
+
+### Tools
+
+LangChain tools for agent use: memory search, recall, and explicit add. Factories take Collection references; no pool:
+
+```python
+from threetears.agent.memory import (
+    load_memory_add_tool,
+    load_memory_search_tool,
+    load_memory_recall_tool,
+)
+
+search_tool = await load_memory_search_tool(
+    user_id=user_id,
+    embedding_provider=embedding_provider,
+    agent_id=agent_id,
+    customer_id=customer_id,
+    authorizer=authorizer_bundle,
+    memories_collection=memories_collection,
+    media_content_collection=media_content_collection,
+    memory_chunk_collection=memory_chunk_collection,
+)
+recall_tool = await load_memory_recall_tool(
+    user_id=user_id,
+    agent_id=agent_id,
+    customer_id=customer_id,
+    authorizer=authorizer_bundle,
+    memories_collection=memories_collection,
+    media_content_collection=media_content_collection,
+    memory_chunk_collection=memory_chunk_collection,
+)
+add_tool = await load_memory_add_tool(
+    user_id=user_id,
+    conversation_id=conv_id,
+    message_id=msg_id,
+    embedding_provider=embedding_provider,
+    agent_id=agent_id,
+    customer_id=customer_id,
+    authorizer=authorizer_bundle,
+    memories_collection=memories_collection,
+)
+```
+
+### Configuration
+
+```python
+from threetears.agent.memory import MemoryConfig
+
+config = MemoryConfig(
+    similarity_threshold=0.4,      # minimum cosine similarity for retrieval
+    detail_threshold=0.85,         # threshold for including full memory detail
+    context_budget=15,             # max memories in context
+    dedup_threshold=0.85,          # similarity threshold for deduplication
+    max_candidates=10,             # max candidates per extraction
+)
+```
+
+## Database Schema
+
+Requires PostgreSQL with the `pgvector` extension. The package's own migration runner (`threetears.agent.memory.migrations.register`) produces the full schema per agent schema. Registered versions:
+
+- **v001** -- `memories` (PK `memory_id`, pgvector `embedding`, scoping ids, content, summary, lifecycle timestamps).
+- **v002** -- `conversation_memory_refs` (ledger of per-conversation surfaced items).
+- **v003** -- column reconciliation: renames PK and discriminator to match the package code (`id` to `memory_id`, `memory_type` to `type_memory`), drops columns the code does not read (`embedding_model`, `importance`, `metadata`, `date_accessed`), loosens `agent_id`/`customer_id` to NULL.
+- **v004** -- lifecycle + conversation-link columns on `memories` (`conversation_id`, `message_id_source`, `is_deleted`, `media_id`, `date_deleted`, `summary`) with indexes.
+- **v005** -- FTS: `search_vector TSVECTOR` + GIN index + maintenance trigger on `memories`.
+- **v006** -- `media` (parent) + `media_content` (chunked extracted text with embedding + FTS).
+- **v007** -- `memory_chunks` (document-style chunks with heading / page metadata + embedding + FTS).
+
+Every FTS column is trigger-maintained from `content` + `summary` (weighted A/B); callers do not have to populate `search_vector` manually. Integration tests under `tests/integration/` exercise the full chain + every public API surface against `pgvector/pgvector:pg16` via testcontainers.
+
+## RBAC Enforcement
+
+Memory reads, writes, and extractions flow through the unified rbac evaluator in `threetears.agent.acl`. Every (agent, customer) pair is a `memory`-type namespace in the `namespaces` table; each access resolves the namespace and evaluates one of three canonical actions against the caller's `(user_id, agent_id)` pair:
+
+- `memory.read` -- retrieval / search / recall. Guarded on `MemoryRetriever.retrieve*`, `MemoriesCollection.find_by_user`, `MemoriesCollection.find_by_scope`, the `memory_search` + `memory_recall` LangChain tools.
+- `memory.write` -- user-initiated writes. Guarded on `MemoriesCollection.save_memory` and the `memory_add` LangChain tool.
+- `memory.extract` -- agent-internal extraction path. Guarded on `MemoryExtractor.extract`; the owner short-circuit keeps the common case (agent emitting memories on its own namespace) grant-free.
+
+Owner short-circuit: the evaluator allows any action when the calling agent owns the memory namespace. Agent-internal retrieval and extraction therefore work without explicit grants; user-initiated reads and writes require evaluator assignments.
+
+Auto-assignment on first user-write: `memory_add` ensures a `MemoryOwner` assignment for the calling user on their first write (idempotent-by-state; the ensurer only fires when the user has zero memory rows in the target schema). Subsequent writes authorize against the materialized grant; admin-revoked grants stay revoked (the ensurer does not resurrect them).
+
+Wiring shape: every consumer of the memory surface REQUIRES a `MemoryAuthorizerDependencies` bundle exposing:
+
+- `acl_cache` -- shared `threetears.agent.acl.AclCache` instance;
+- `membership_loader` + `grant_loader` -- the evaluator's loaders (`threetears.agent.acl.MembershipLoader` / `GrantLoader`);
+- `namespace_collection` -- three-tier `NamespaceCollection` used to resolve the memory namespace via `get_by_owner_and_customer(namespace_type="memory", owner_agent_id, customer_id)` (create-if-absent flows through `save_entity`);
+- `group_collection` + `group_member_collection` + `role_collection` + `role_assignment_collection` -- the rbac Collections the first-write owner-assignment path uses via `ensure_memory_owner_assignment(...)`.
+
+There is no bypass. Every `MemoriesCollection`, `MemoryRetriever`, `MemoryExtractor`, and LangChain tool factory (`load_memory_search_tool`, `load_memory_add_tool`, `load_memory_recall_tool`) takes the bundle as a required constructor/factory argument; every code path that touches a memory row runs `authorize_memory_access` first. Callers that omit the bundle fail at the type checker and the Python signature boundary.
+
+- Production wiring builds the bundle directly from the agent-side three-tier stack's Collections (`NatsProxyL3Backend`-backed `NamespaceCollection` / `GroupCollection` / ...).
+- Test wiring injects a permissive fixture `permissive_memory_authorizer` (see `tests/conftest.py`) that carries in-memory Collection stand-ins and a permissive evaluator. Fixture usage is explicit in every test file that constructs a memory surface.
+- Back-office / admin tooling that genuinely needs to read or write memories without an identity must construct its own bundle with Collections bound directly to an asyncpg pool; there is no global escape hatch.
+
+See `threetears.agent.memory.authorize` for the full public surface.
+
+The three platform roles (`MemoryOwner` / `MemoryReader` / `MemoryWriter`) carry the canonical action vocabulary. Platform-side migrations seed these roles and backfill the rbac rows required for evaluator resolution.