Project File

D-CSIL Platform — Cortex & Harness

Durable shared memory (Cortex) and disciplined session orchestration (Harness) for every agent in the lab, served from a Raspberry Pi 5 named Bailey.

OperationalOne memory, many agents.

Cortex

MEASURED

A local-first memory appliance: 20 memory types across 7 scopes, hybrid retrieval (FTS5/BM25 keyword + vector similarity + recency + confidence weighting), automatic dedupe, TTL and supersede chains, secret scanning that rejects credential-shaped writes, budgeted context packs, and an append-only JSONL event log. It serves REST and exposes 12 MCP tools. It never runs model inference. 41 tests pass.

Harness

MEASURED

The Mac-side runtime that runs agent sessions — Claude Code, Codex, Fable, OpenCode, Grok Build — against Cortex with a start → context pack → agent → writeback → close lifecycle. It parses each agent's final report and distills it into memory, queues writebacks to disk when Bailey is unreachable, and serves a FastAPI dashboard. Fable orchestration mode decomposes tasks into tiered plans routed across cheap/mid/premium models. 23 tests pass, including a live cross-repo Cortex contract test.

Current Limitations

−Cortex's graph layer is a stub (NullGraphBackend); weighted-graph retrieval is v2 work.
−The commercial/hosted Cortex site is a web shell only — the engine is the local appliance.

Next Steps

+V2 index-first retrieval path for large knowledge bases (non-destructive rebuild in progress).
+Obsidian vault view as a first-class read surface.