Agent Memory Kits
agent-memory-kitsHost-specific setup, lifecycle integration, doctors, ingestion, recall, compaction, and claim/evidence workflows.
Choose a host →RESOLVING ROUTEFollowing the evidence boundary…
Give the coding agent you already use durable, local-first memory—then inspect what it retrieved, what changed over time, and what the evidence can actually support.
Repository pushes and crate releases update from public sources. If an upstream service fails, the interface says so and falls back to a dated snapshot instead of pretending freshness.
Adopt the open-source memory plane, evaluate its evidence model, or bring the architecture into a real integration boundary.
Choose your host, run a short verified path, and keep authority on your machine.
Open the Install Cockpit →02 / EVALUATEWalk through a redacted receipt, replay boundary, and the claims the system refuses to make.
Enter the Proofroom →03 / INTEGRATEScope an agent-memory integration or evidence/release-truth audit with Josh Stevenson.
Explore integration services →Choose the layer that matches the job: operator kits, a protocol server, or the authoritative Rust retrieval core.
agent-memory-kitsHost-specific setup, lifecycle integration, doctors, ingestion, recall, compaction, and claim/evidence workflows.
Choose a host →semantic-memory-mcp · 0.5.4Bounded MCP profiles, witnessed retrieval, replay, graph access, and separate assertion/action authority decisions.
Inspect the server ↗semantic-memory · 0.5.11SQLite-authoritative hybrid retrieval, weighted RRF, bitemporal views, explained ranking, and durable receipts.
Inspect the core ↗The retrieval path stays inspectable. Degradation in one plane cannot silently promote another.
state CURRENT · receipt mcp-witness-… · replay inputs NOT RETAINEDReceipt ≠ truthSix responsibilities form the trust substrate. Open a plane for its packages, dependency shape, and explicit limitations.
Durable retrieval, temporal state, claims, and source lineage.
7 packages →02Permits, decisions, attestations, settlement, and refutation.
22 packages →03Wire-visible schemas, canonical IDs, and boundary compilation.
5 packages →04Typed graphs, bounded plans, resumability, and tool effects.
7 packages →05Measured vector and KV codecs with explicit quality gates.
9 packages →06Fail-closed admission, sandboxing, policies, and typed failure.
7 packages →Maturity labels describe sequencing—not blanket readiness claims.
Memory, tool calls, policy decisions, and final claims share one inspectable evidence graph.
semantic-memory · llm-tool-runtime · agent-graph · claim-ledgerA release boundary that refuses stale status, unsupported claims, and unbound artifacts.
claim-ledger · attestation-exchange · verification-policy · receipt-benchLocal-first working memory that understands what changed, what was superseded, and what still needs proof.
semantic-memory · bitemporal-runtime · living-memory · knowledge-runtimeStrong subsystem evidence exists. That does not automatically certify a release, prove correctness, or authorize an action.
Source, package, live, and receipt evidence remain separate scopes.
Start with the guided open-source path. If the integration touches release truth, governance, or evidence architecture, scope the boundary directly with Josh.