Scope Recall is a local memory core for coding and assistant agents, with SQLite as the only fact authority and rebuildable vector indexes alongside it. It plugs into Hermes as a memory provider, and into Codex, Claude Code, WorkBuddy, and DeepSeek Harness through a shared adapter of hooks and an MCP server.
Several agents can attach to one shared store: what the owner tells one agent another can recall, each memory records which agent it came in through, and a deletion through any of them applies to all. Memory accumulates evidence before a fact is written rather than judging a single sentence on sight, and tool output is kept and searchable but never turned into facts.
Read views return a categorized current-fact profile or the statements about one entity, built only from admitted consolidated claims, and a trace operation joins two or three recorded relationships into a bounded multi-hop evidence path without calling another model. Installation, upgrades, and migrations from older databases are designed to be run by the agent itself through a bundled setup skill, and a second skill answers the owner's questions about what is remembered and how to correct, mute, or delete a memory.
The release runs about 2,300 tests that check contracts, storage, and host wiring, and none measures recall quality: accuracy figures come from the maintainers' own hand measurements, and there is no automated regression suite for them.
Features
- SQLite fact authority: facts live in SQLite, with vector indexes as rebuildable companions
- Host adapters: Hermes memory provider, plus hooks and an MCP server for Codex, Claude Code, WorkBuddy, and DeepSeek Harness
- Shared store: several agents attach as entries to one store, with per-memory source tracking and deletions that apply everywhere
- Evidence accumulation: facts are written only after evidence builds up, and tool output is never promoted to facts
- Profile and entity views: deterministic read-only views over admitted claims, within a token budget
- Multi-hop trace: joins two or three recorded relationships with node, path, time, and byte budgets
- Hybrid retrieval: semantic search through an optional LanceDB companion
- Agent-operated setup: a setup skill routes fresh installs and legacy migrations, including backup, indexing, and host checks
- Memory governance skill: answers what is remembered about the owner and what correcting, muting, or deleting a memory does
- Receipt-driven uninstall: removes plugin files and keeps memory data unless a purge is planned and applied separately
