Summary
Talos Memory packages a versioned memory core and a separate policy registry. Applications can trace a derived result to its source versions, check whether that evidence is still usable, and assemble typed context within an explicit budget. The MIT-licensed alpha runs on Linux without a model or external database.
Role and scope
I design the storage and retrieval contracts, implement the Python libraries and interfaces, and build synthetic tests for failure, concurrency, withdrawal and recovery. I extracted the reusable components into a dedicated public repository with bilingual documentation and standalone examples.
Problem and constraints
A search index or restored backup can still contain evidence that was corrected or withdrawn. A useful memory system must distinguish current evidence, outdated versions and unavailable validation before returning context to an application.
Architecture
SQLite stores immutable versions and explicit dependency references. A separate authority and freshness witness govern admissibility; disposable FTS5 indexes propose candidates that are revalidated on read. Typed envelopes distinguish documents, observations, hypotheses, procedures, tasks and feedback. Python, CLI, optional HTTP and MCP interfaces expose the contracts.
Decisions and trade-offs
The index is a projection, not the authority. Writes use compare-and-swap; recovery from an old content snapshot begins in quarantine against a retained current authority. The optional registry separates policy proposal, evidence, admission and rollback, without activating policies or granting permissions.
Testing and verification
The public release contains 122 memory tests and 36 policy-registry tests. CI passed on Python 3.11 through 3.14. Released distributions were also downloaded anonymously, installed outside the source tree and checked with all 158 tests and synthetic demos on Linux.
Public results
Version v0.2.0a1 is available with wheels, source distributions, checksums and EN/FR documentation. The two-actor demo withdraws one source, denies its dependent result, preserves an unrelated source and verifies recovery from an old content backup. These checks demonstrate library contracts on synthetic data.
Resources and links
Scope and current status
The frozen R3 retrieval policy improved evidence selection in a bounded synthetic comparison but missed its relative latency gate. No production speed gain is claimed. Recovery requires a current authority and freshness witness outside the restored content snapshot; rolling all three back together is outside the guarantee. Revocation controls future library reads, not copies already exported or model unlearning.