Skip to content

ROADMAP.md

Latest commit

 

History

History
86 lines (71 loc) · 4.7 KB

ROADMAP.md

File metadata and controls

86 lines (71 loc) · 4.7 KB
Raw

Roadmap

What mcp-recall is aiming at, what it is deliberately not becoming, and where the two open long-horizon ideas sit. This is a statement of direction, not a dated plan — a solo open-source project makes no delivery promises. For detailed per-phase status, see the phase table in CLAUDE.md; for the shape of the system, docs/architecture.md.

The bet

mcp-recall occupies one specific gap: layer ③ of the context stack (see the README). Claude Code's native microcompaction offloads built-in tool output, but MCP tool output is truncated at a 25k-token ceiling and discarded. mcp-recall intercepts MCP (and Bash) output before it reaches the window, stores the full payload locally, and keeps it retrievable by full-text search across sessions.

Everything below is judged against that bet. A change that sharpens the MCP-output-capture-and-retrieval story is in scope; a change that turns the project into a different kind of tool is not — however useful that other tool might be.

Directions (themes, not commitments)

These are the areas active work tends to land in. No dates, no ordering promises — they describe where effort goes when it goes somewhere.

  • Retrieval quality — making Claude's first recall land on the right content: better summary hints, graduated/peek retrieval, structure-aware fallbacks. The store is only as good as the odds a single query finds what's needed.
  • Robustness & store lifecycle — the store must be safe to leave running for months: correct eviction, disk reclamation (gc), dedup, and no silent data-loss edges. Correctness of the delete path outranks any feature.
  • Contributor experience & docs accuracy — handlers and profiles should be cheap to add and the docs should never overstate what the code does. Phase 13 exists entirely for this.
  • Profile ecosystem — the community-profile tier is how coverage scales without a TypeScript change per MCP; lowering the cost of contributing and trusting a profile stays a throughline.

On the horizon (deferred, not committed)

Two filed ideas are genuine extensions of the bet but are not being built now. They are deferred, not rejected — if they ship, they ship as described here.

  • Anthropic memory-tool backend (#188) — expose mcp-recall's verbatim + FTS store as a backend behind the API memory tool. A natural fit: the memory tool is a model-managed summary scratchpad with no verbatim retrieval, which is exactly what mcp-recall already provides. Deferred on effort, not on scope.
  • Optional local hybrid FTS + embeddings (#189) — semantic recall alongside the deterministic FTS index. Opt-in only, never a default requirement — the default retrieval path stays zero-heavy-dependency (see non-goal 4). If it lands, it is a flag, not a baseline.

Non-goals

The boundaries that keep the project focused. Each is a door deliberately left closed; reopening one is a direction change, not a feature request.

  1. Not conversational or semantic memory. mcp-recall stores raw tool outputs verbatim — it does not distill a model of the conversation, build a knowledge graph, or summarize what happened. That is a different product (mem0, Letta, Zep, and similar occupy it). Keeping to raw-output capture is what makes verbatim retrieval and secret-safe storage tractable.
  2. Won't intercept the built-in Read / Grep / Glob tools. Claude Code's PostToolUse output-replacement only supports MCP tools and Bash; the other built-ins can't be intercepted, and native microcompaction already offloads them. Scope is MCP + Bash, by platform and by design — mcp-recall does not try to beat the native path for built-ins.
  3. No cloud, hosted service, telemetry, or accounts. Local-first, per-project SQLite, no server to run and no phone-home. There is no plan for a sync backend or a hosted tier; your stored output stays on your disk.
  4. The default retrieval path stays zero-heavy-dependency. Deterministic SQLite FTS5, no required embeddings model or ML runtime to install or run. Any semantic-search support (see #189) is opt-in and additive — it never becomes a precondition for using mcp-recall.

Proposing a change of direction

A non-goal is not permanent, but changing one is a deliberate decision, not a drive-by PR. Open an issue that names the non-goal and argues why the bet has changed. Everything else — new handlers, profiles, retrieval and lifecycle improvements — is welcome without ceremony; see CONTRIBUTING.md.