reconcile-agent-memory 0.1.0

A pluggable memory layer for LLM agents with usage verification, contradiction resolution, and multi-tenant access control

Reconcile

A pluggable memory layer for LLM agents that fixes a gap in every major alternative (Mem0, Letta/MemGPT, Zep): none of them verify that a fact retrieved and injected into an agent's context was actually used. This does.

What it does

• Write-time contradiction resolution — a new fact about the same (tenant, subject, predicate) supersedes the old one instead of both floating in the store forever. O(1) via an index, not a linear scan.
• Hybrid retrieval — TF-IDF similarity + subject-match bonus, scoped per tenant.
• Usage verification — after a fact is injected into an agent's context, checks whether the response actually reflects it (not just whether the right words appear, but whether they're framed as current state vs. a suggestion, with an optional semantic-similarity fallback for paraphrases).
• Multi-tenant access control — TenantScopedStore enforces caller→tenant authorization, not just a partition key.
• Rule-based extraction — a regex fallback for turning raw text into facts when no LLM is available (explicitly not a replacement for one).
• Thread-safe writes — a lock around the contradiction-check-and-write.

Status

65/65 tests passing. This is a validated prototype, not a production system — every known limitation is documented in STRESS_TEST_SUMMARY.md, SCALE_NOTES.md, and PARAPHRASE_CALIBRATION.md, and several tests exist specifically to pin known gaps rather than hide them (search test files for known_gap / known_limitation).

Install