Skip to content

Memory layer setup: claude-flow MCP + Phase 3 design handoff#2

Open
aryansingh-rgb wants to merge 1 commit into
mainfrom
smash-memory-mgmnt
Open

Memory layer setup: claude-flow MCP + Phase 3 design handoff#2
aryansingh-rgb wants to merge 1 commit into
mainfrom
smash-memory-mgmnt

Conversation

@aryansingh-rgb

Copy link
Copy Markdown
Collaborator

No production Go code in this PR. Implementation lands in a follow-up branch per HANDOFF.md §6/§7.

Why design-only

A live MCP probe this session revealed the tool surface differs materially from earlier handoff guesses (memory_retrieve not memory_get, similarity not score, 768-dim not 384-dim, search returns truncated value). Capturing real shapes before writing Go bindings saves the next dev meaningful debug time. The probe also showed our placeholder threshold (0.95) is far too high — a near-paraphrase scored only 0.6366 — so calibration is parked in §13.1 and runs alongside Phase 3 implementation rather than blocking this PR.

Test plan

  • Reviewer reads HANDOFF.md §0–§7 and §13–§16
  • Teammate clones the branch, opens Claude Code in the directory, accepts the project-scope MCP prompt, confirms claude mcp list shows claude-flow ✓ Connected
  • Teammate runs the threshold calibration battery (§13.1) before locking the value
  • No CI surface changes; existing CI passes unchanged

Adds project-scope claude-flow MCP server config and a detailed
session-bridging handoff document. No production Go code yet --
implementation is the next teammate task per HANDOFF.md sections 6/7.

What is included:
- .mcp.json: project-scope claude-flow@alpha MCP server (memory_store,
  memory_retrieve, memory_search exposed; HNSW + sql.js backend with
  768-dim mpnet embeddings).
- HANDOFF.md: 600+ line bridging doc with locked design (short-circuit
  only memory pattern, threshold in config, resolver-side firing),
  live-probed real MCP tool surface (replaces earlier guesses), Phase 3
  Go implementation plan with full code shape, threshold calibration
  battery (section 13.1) the next session must run before locking the
  value.
- .gitignore: ignore .swarm/ (claude-flow runtime SQLite state).

Key probe finding: near-paraphrase scored 0.6366 against this embedding
model -- the placeholder 0.95 short-circuit threshold would never fire.
Calibration deferred to teammate per section 13.1.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant