Pilot: can one reviewed project gotcha survive a later coding-agent session?

[zzhang82]

I maintain Agent Memory Bridge. I am looking for concrete reports from developers who use coding agents and want project gotchas to survive between sessions.

The current reference point is v0.22.1: a local-first stdio MCP server with 10 public MCP tools and a release validation snapshot of 395 passed. Those are local project checks, not external adoption evidence.

This pilot starts with one client. The cross-client receipt is an optional second phase.

Phase 1: One Client, Two Sessions

The question is simple: can one reviewed, non-sensitive project gotcha be stored in one session and recalled in a later session?

1. Install AMB and configure one MCP client to use a local AMB home.
2. Choose one reviewed project gotcha that contains no secrets or proprietary information.
3. In the first session, use the MCP store tool:

store(
  namespace="project:<repo>",
  kind="memory",
  title="Reviewed project gotcha",
  content="record_type: gotcha\nclaim: Run the generator after schema edits.",
  tags=["reviewed:true"],
  source_client="<caller-declared label>"
)

4. End that session. In a later session using the same configured client and AMB home, use:

recall(
  namespace="project:<repo>",
  query="generator after schema edits"
)

5. Report whether the later session found the same gotcha, what was unclear, and how long the attempt took. Timing is measured, not promised.

Optional Phase 2: Two Clients

If Phase 1 works and you already use two MCP clients on the same repository:

1. Configure client B against the same AMB home.
2. Use one namespace and correlation ID across both clients.
3. Have client A call store(...) for one reviewed writer memory.
4. Have client B call recall(...), then store(kind="signal", ...) and ack_signal(...).
5. Run the local activation-receipt report.

See the cross-client instructions and receipt boundaries.

What I Want to Learn

• Did a later session recall the same gotcha?
• Which client and OS did you use?
• Which install, configuration, or tool-call step was unclear?
• Was the recalled record useful compared with native memory or repository instruction files?
• If you tried Phase 2, did the receipt pass, require review, or fail?

Failed attempts are useful. Please remove secrets, proprietary content, private paths, session IDs, and model IDs from reports.

Boundaries

• AMB does not automatically turn transcripts into durable memory.
• Durable records should be reviewed before they are stored.
• source_client is caller-declared provenance. It is not authenticated identity.
• The local activation receipt is not certification, vendor support, or external adoption proof.
• The receipt does not prove that the stored memory is factually correct.
• The pilot makes no completion-time promise.

Start Here

• v0.22.1 release
• Agent install guide
• Client integrations
• Client provenance and receipt boundaries

Reply Template

Client(s):
OS:
Phase 1 timing:
Later-session recall: yes / no
Recalled record useful: yes / no / unsure
Hardest or unclear step:
Phase 2 tried: yes / no
Receipt: pass / review-required / failed / not tried
Private details removed: yes / no