Documentation

Start with the agent-facing path

01 / STARTInstall and initialize

Install one binary, initialize a project, and inspect the ambient context your agent will inherit.

02 / RECALLRetrieve one bounded packet

Receive qualified claims or exact, cited recovery evidence without changing honest absence semantics.

03 / CONTEXTGive an agent bounded context

Build a citation-rich packet within a byte budget while keeping retrieved content explicitly untrusted.

04 / PRESERVEPreserve a deliberate checkpoint

Stage a small handoff note outside recall, then review and promote it through the normal grounding boundary.

The shortest useful workflow

cd /path/to/project
memoree init --name my-project
memoree retrieve "what do we know about the deployment decision?"
memoree context build "deployment constraints" --max-bytes 4096
memoree remember --apply "Production deploys require a clean migration check."

Normal retrieval stays within ambient project/task context. Pin the target repository and run every ambient command there; a none result means no qualified match at that horizon, not permission to widen the search. Recovery bytes are exact and cited but remain unqualified. Use repository tools for current code; repository state overrides dated memory. Broaden to workspace or personal scope only for a separate request with a reason. The project index is an explicit experimental evaluation surface, not part of the canonical agent route.

Understand the boundaries

• Architecture explains local storage, immutable revisions, scope, and why indexes remain disposable.
• Context and configuration defines markers, task-local sessions, personal fallback, and application paths.
• Project source indexing documents the experimental bounded structural map, its negative downstream result, and guarded reindexing.
• Project metrics and experiments explains opt-in real-operation traces and honest randomized token/time comparisons.
• Quality gates records the claims the project is willing to make and the evidence required before making stronger ones.
• Retrieval evaluation separates deterministic correctness checks from downstream model quality.