Working product · Trusted agent memory
AI agents retain useful context without mistaking memory for current truth.
MemCore is a local-first memory and trust layer that captures scoped context, retrieves relevant history, builds a structured world model, and separates ordinary recall from source-backed current truth.
The problem
Agents forget between sessions, but saving everything creates a second problem: noise, stale facts, cross-project leakage, contradictions, and confident answers built on memory that may no longer be true.
What we built
AI agents retain useful context without mistaking memory for current truth.
Scoped capture, semantic and keyword recall, entity and relationship memory, lessons and procedures, lifecycle maintenance, provenance, a Claim Ledger, Trust Kernel, resolver APIs and MCP, and trusted project-state synthesis.
What changed
Agents regain the right history before a prompt while stale, conflicting, or weakly supported claims can be marked as contextual, last-known, or needing verification instead of being presented as current truth.
How it runs
MemCore captures useful facts, recalls through multiple retrieval paths, ranks context, detects contradictions, and synthesises current project state only from trusted claims.
Operators set scopes and approved sources, control trust and retention policy, review high-risk exceptions, and can inspect provenance, export, or delete scoped memory.
Evidence
Working product, HTTP and MCP resolver surfaces, automated recall evaluations, privacy and tenant-isolation gates, and a content-free Evidence Command Center.
Product evidence only; no customer outcome is claimed. Private memory contents, scopes, runtime data, and client identities remain private.