What is BrainCore?
An open-source operational memory system for AI agents that preserves artifacts, facts, evidence, time, and trust status.
BRAINCORE
Evidence-grounded operational memory for AI agents.
BrainCore preserves operational knowledge as artifacts, evidence-linked facts, trust classes, and temporal retrieval so agents can recover what changed, what fixed an issue, and which claims are supported by source evidence.
DIRECT_ANSWER
BrainCore is operational memory, not generic chat memory.
BrainCore is designed for long-running AI infrastructure work where agents need context that survives handoffs. It preserves source artifacts first, extracts evidence-linked facts, assigns trust status, and retrieves context through PostgreSQL, full-text search, pgvector similarity, temporal expansion, and optional graph-path retrieval.
PUBLIC_PROOF
Production-corpus proof points are useful, but scoped.
| Proof Point | Public Claim | Boundary |
|---|---|---|
| 26,966 facts | BrainCore has committed production-corpus artifacts reporting 26,966 extracted facts. | This is a corpus measurement, not a universal quality claim. |
| 71.6 ms P50 / 85.2 ms P95 | Benchmarks report retrieval latency with vector stream disabled. | Do not describe this as benchmark-leading or representative of every deployment. |
| 98.52% fact-evidence coverage | Claims verifier artifacts report high evidence coverage for extracted facts. | Evidence coverage is not the same as SOTA retrieval relevance. |
| v1.1.6 | The latest public release is documented as v1.1.6 as of April 30, 2026. | Describe BrainCore as early open source, not a hardened remote enterprise appliance. |
ARCHITECTURE
Archive, extract, ground, retrieve.
BrainCore's architecture is built around inspectability. Source artifacts remain visible, facts are linked to evidence segments, and retrieval can use structured SQL, full-text search, vector similarity, time-aware expansion, and graph-path context.
Source artifacts first
Incidents, session notes, benchmark outputs, and operational records remain traceable.
Evidence-linked facts
Extracted memory points back to supporting evidence instead of standing alone as opaque summaries.
PostgreSQL plus pgvector
Structured data, SQL, full-text search, and semantic retrieval live in one inspectable memory layer.
FIT
Who BrainCore is for.
BrainCore is for builders of long-running AI systems that need operational memory: agent runtimes that need project continuity, MCP clients that need grounded retrieval, SRE and platform teams preserving incident lessons, and local-first AI builders who want Postgres-native memory. It should not be positioned as a complete incident management suite, a generic chat memory plugin, or a finished remote MCP service.
Public-claim boundary: BrainCore's local memory stack and benchmark artifacts are public proof. Remote deployment hardening, tenant isolation policy, and enterprise adoption require separate evidence before public claims.
FAQ
BrainCore questions.
How is BrainCore different from RAG?
RAG retrieves chunks. BrainCore also tracks source artifacts, extracted facts, provenance, temporal context, trust classes, and playbook candidates.
Is BrainCore production-ready?
It is early open source with committed benchmark artifacts and architecture docs. Remote MCP and security wrappers require operator hardening.
What can agents ask BrainCore?
What changed, what fixed this last time, what evidence supports that, which facts are trusted, and what playbook emerged.
RELATED_TOPICS