Agent Memory System

Agent Memory System

[](https://github.com/rrrrrredy/agent-memory-system/actions/workflows/ci.yml) [](LICENSE)

Local-first, evidence-backed memory for coding agents.

Agent Memory System preserves the task evidence a runtime exposes, rebuilds it into episodes, turns supported observations into reviewable memory candidates, and shares only redacted memories with explicit validation and promotion records through a separate private Git repository. Deployment policy requires an operator to make those decisions; the software binds each attestation to the evidence but does not authenticate that the caller is human.

It is built for a harder question than “what should the agent remember?”:

What evidence supports this memory, is it still current, and can another machine use it without receiving the private transcript?

Version **0.3.0** adds immutable memory loadouts, evidence-bound review packets, replay-verifiable local Codex execution receipts, prospective longitudinal study records, and a loopback-only read-only dashboard to the storage, derivation, review, private Git, retrieval, recovery, and evaluation foundation. Runtime adapters remain version-sensitive. The system fails closed when evidence is incomplete or a compatibility claim cannot be verified.

Why this is different

Most memory tools summarize a conversation and place the summary in a retrieval index. This project separates evidence, judgment, and distribution:

flowchart LR
    A["Agent-local traces"] --> B["Append-only evidence ledger"]
    B --> C["Episodes and compaction checks"]
    C --> D["Candidate memories"]
    D --> E["Validation attestation"]
    E --> F["Promote, supersede, or revoke"]
    F --> G["Private Git memory repository"]
    G --> H["Bounded cross-agent retrieval"]
    H --> I["Immutable memory loadouts"]
    I