v0.5.0 - Measurable retrieval and fuller signal lifecycle

Agent Memory Bridge v0.5.0

Agent Memory Bridge is a two-channel MCP memory layer for coding agents:

• durable knowledge in memory
• coordination events in signal

This release turns two important parts of the bridge into something we can measure and trust.

Retrieval became measurable and tunable

• added a canonical benchmark fixture for coding-memory retrieval
• benchmark reports now track precision@1, precision@3, and expected_top1_accuracy
• retrieval now reranks a larger candidate pool instead of relying on raw FTS order
• the benchmark compares bridge recall against a simple file-scan baseline
• the current canonical fixture reports:
  • memory_expected_top1_accuracy = 1.0
  • file_scan_expected_top1_accuracy = 0.5

Signal lifecycle became more production-shaped

• signal now supports claim -> extend -> ack / expire / reclaim
• added extend_signal_lease to the MCP surface
• lease renewal is separate from reclaim
• stale leases must be reclaimed instead of extended
• hard signal expiry remains a strict upper bound for any extension

Proof and regression coverage

• deterministic proof covers signal correctness, duplicate suppression, and recall timing
• stdio integration covers the MCP tool surface, including lease extension
• full test suite passes: 68 passed

Why this release matters

This is the point where the bridge starts to feel less like a clever memory tool and more like a small coordination-aware framework:

• retrieval quality is now benchmarked instead of guessed
• signal coordination has a cleaner lifecycle boundary
• the public surface stays small and inspectable

Current MCP surface

• store, recall
• browse, stats
• forget, promote
• claim_signal, extend_signal_lease, ack_signal
• export

Roadmap

Next likely steps are:

• claim selection fairness
• minimal retry boundaries for signals
• broader benchmark fixtures beyond the current canonical set
• continued learning-quality work in promotion and synthesis

v0.4.1 - Neutral sample profile and landing cleanup

Agent Memory Bridge v0.4.1

This is a small cleanup release focused on framework adoption and landing consistency.

What changed

• replaced the last Cole-shaped sample profile path with a neutral default profile source root
• updated the core fallback profile source path to use a neutral bridge-local location instead of a personal profile name
• aligned the README setup section with the actual config shape
• added a small benchmark plan stub so the proof track has an explicit starting point
• added a path-level test for the neutral default profile source root

Why it matters

v0.4.0 made the framework shape real.

v0.4.1 makes the first-run experience look more like a reusable framework and less like a private workflow repo.

Verification

• full test suite passes: 58 passed

GitHub: https://github.com/zzhang82/Agent-Memory-Bridge

v0.4.0 - Frameworkization and signal lifecycle

Agent Memory Bridge v0.4.0

This release turns the bridge into a cleaner framework and gives signal a real coordination lifecycle.

What changed

Neutral core defaults

• core defaults now use neutral profile settings instead of Cole-shaped titles, actors, and namespaces
• profile-specific wording now sits in config instead of leaking into the framework runtime
• the default shared namespace is now global

Signal lifecycle

• signal is no longer just append-only state
• added MCP support for:
  • claim_signal
  • ack_signal
• signals now support:
  • lease ownership
  • TTL / expiry
  • visible statuses: pending, claimed, acked, expired

Storage cleanup

• signal lifecycle helpers now live in a dedicated module
• export rendering now lives in its own module
• storage behavior now reflects the two-channel model more clearly

Docs and setup

• README and README.zh-CN were rewritten around the v0.4 shape
• setup now explains the neutral profile config
• roadmap now reflects the next stage: proof, selection, and composition

MCP surface

The public MCP tools are now:

• store
• recall
• browse
• stats
• forget
• promote
• claim_signal
• ack_signal
• export

Verification

• full test suite: 57 passed

Why this release matters

Agent Memory Bridge is not trying to be a general memory platform.

It is a small MCP-native bridge for coding-agent workflows, with two separate channels:

• durable knowledge in memory
• coordination events in signal

v0.4.0 makes that split much more real in both the runtime and the public surface.

v0.3.0 — Two-channel memory for coding agents

Agent Memory Bridge v0.3.0

v0.3.0 makes the bridge easier to trust in day-to-day use. You can inspect what is stored, correct bad entries from a live session, export data back out, and check whether the watcher is still parsing Codex rollout files correctly.

Built for Codex-first workflows.

Two channels, named plainly

Agent Memory Bridge keeps durable knowledge and coordination events separate:

• memory for learned decisions, gotchas, and domain notes
• signal for handoffs, polling, and agent coordination

New MCP tools

• browse(namespace, ...) lists entries when you do not know the right search term yet
• stats(namespace) shows totals, kind breakdowns, domain counts, and oldest/newest entries
• forget(id) removes bad or stale entries
• promote(id, to_kind) reclassifies an entry into learn, gotcha, or domain-note
• export(namespace, format, ...) writes entries back out as Markdown, JSON, or plain text

Runtime checks

• watcher health now verifies that Codex rollout files still parse into usable summaries
• the main healthcheck includes watcher parsing status
• Docker support and Glama metadata are in the repo for MCP validation workflows

Docs and workflow

• the README now leads with the two-channel model and a shorter quickstart
• signal handoff examples and namespace guidance are easier to find
• the roadmap now focuses on memory lifecycle, retention, recall assembly, and evaluation

Verification

• full test suite: 53 passed

What is next

• durable-event scoring
• richer gotcha extraction
• recall assembly policy
• lightweight relationship metadata
• evaluation for recall and promotion quality

v0.2.0 — Codex-first memory shaping for coding agents

Agent Memory Bridge v0.2.0

Agent Memory Bridge is an MCP-native, local-first memory layer for coding agents.

It turns coding sessions into reusable engineering memory:

• summaries
• learned decisions
• gotchas
• domain notes

Built for Codex-first workflows.

What's new in 0.2.0

Product framing

• unified naming around agent-memory-bridge
• removed the legacy cole_mem_bridge compatibility package
• cleaned up the public repo surface and documentation
• added a clearer positioning note for how this differs from broader memory platforms

Codex workflow support

• agentMemoryBridge is the canonical MCP server
• startup protocol now explicitly includes project:<workspace> when a workspace exists
• session visibility improved with session-seen
• automatic checkpoint and closeout flows remain in place

Memory shaping

• sessions promote into:
  • summary
  • learn
  • gotcha
  • domain-note
• first-pass domain consolidation is now part of the runtime
• memory remains machine-first and token-aware instead of transcript-heavy

Runtime cleanup

• fixed startup launcher cleanup
• smoke-tested MCP behavior after restart
• re-verified the watcher and service chain
• tests passing: 44 passed

What this project is for

Agent Memory Bridge is intentionally narrow.

It is built for:

• coding-agent workflows
• MCP-native integration
• local-first deployment
• reusable engineering memory across sessions and projects

It is not trying to be a universal cognitive memory platform.

Core idea

session -> summary -> learn -> gotcha -> domain-note

Status

The foundation is in place:

• MCP autoload in Codex
• project and session sync
• recall-first workflows
• reflex promotion
• first-pass domain consolidation

Roadmap

Next areas of work include:

• stronger durable-event scoring
• better gotcha extraction
• richer domain/topic synthesis
• later worker/task-query support

See the roadmap in docs/ROADMAP.md.