PRODUCTION-STATUS.md

Production Status

Last updated: 2026-04-19 (America/New_York)

This maintainer note describes the released 0.9.0 shape plus the validation snapshot used to support it.

Released 0.9.0 Runtime Shape

agent-memory-bridge now has these cooperating layers:

1. stdio MCP server for store, recall, browse, stats, forget, promote, claim_signal, extend_signal_lease, ack_signal, and export
2. shared SQLite/WAL + FTS5 bridge storage
3. session watcher and checkpoint/closeout capture
4. reflex promotion into machine-first durable artifacts
5. consolidation with compression-aware domain-note, belief-candidate, and belief generation
6. relation-lite metadata parsing and surfacing
7. profile/control-layer startup assembly
8. local metadata-only telemetry
9. concept-note emission from stable beliefs
10. task-time assembly over procedures, concepts, beliefs, and linked supporting records

Verified On 2026-04-19

• pytest passes: 152 passed
• deterministic proof reports 4/4 checks passed
• deterministic proof and benchmark both report relation_metadata_passed = true
• benchmark summary reports:
  • question_count = 11
  • memory_expected_top1_accuracy = 1.0
  • memory_mrr = 1.0
  • file_scan_expected_top1_accuracy = 0.636
  • file_scan_mrr = 0.909
  • duplicate_suppression_rate = 1.0
• reviewed classifier calibration snapshot reports:
  • sample_count = 16
  • classifier_exact_match_rate = 0.875
  • fallback_exact_match_rate = 0.062
  • classifier_better_count = 13
  • fallback_better_count = 2
  • classifier_filtered_low_confidence_count = 2
• healthcheck includes a relation-metadata smoke path
• relation-lite metadata is available on recall, export, and stats for:
  • supports
  • contradicts
  • supersedes
  • depends_on
  • valid_from
  • valid_until
• promotion re-derives relation and validity tags after content rewrite, preventing stale derived-tag drift
• consolidation emits a first-class concept-note once a stable belief is promoted
• task-time memory assembly composes:
  • kind:procedure
  • kind:concept-note
  • kind:belief
  • linked supporting records via relation metadata
• recall_first(...) can surface procedure, concept, belief, and supporting layers alongside project/global gotcha and domain retrieval

What 0.9.0 Actually Means

• the public MCP surface is still the same small bridge
• relation-lite structure is now real and auditable
• retrieval claims are benchmarked instead of guessed
• the engine can assemble task-time memory over procedures, concepts, beliefs, and linked support without exposing a larger MCP API

Honest Boundaries

The release still does not mean:

• a graph database
• full relation-aware traversal or ranking across the whole store
• automatic procedure learning from raw transcripts
• a full agent runtime or scheduler
• pre-compaction capture before model-side context loss
• active pubsub or consumer execution on top of stored signals

Pressure Points After 0.9.0

The most important remaining gaps are:

1. pre-compaction capture before model-side loss
2. broader reviewed retrieval fixtures so credibility does not overfit the current corpus
3. stronger write-side calibration for promotion quality
4. cross-domain concept synthesis beyond the current domain-local concept-note step
5. more deliberate procedure curation or promotion instead of only manual procedure records
6. richer coordination semantics only after the memory/governance lane is fully stable

Maintainer Read

0.9.0 is the point where the project reads as more than foundation work.

It now behaves like:

• a shared MCP memory backend
• a governed learning layer
• a structured relation-lite memory layer
• a first pass at applicable/compositional task memory

The next work should protect those gains instead of widening the public surface too fast.