v0.14.0 - Governed Learning Candidates

v0.14.0 - Governed Learning Candidates

Agent Memory Bridge 0.14.0 adds a governed staging lane for runtime learning candidates while keeping the public MCP surface small and stable.

Thesis

0.14.0 = policy-gated learning candidates before durable memory writeback.

This release adds the missing boundary between a runtime noticing something useful and AMB treating it as ordinary durable memory. Runtime or harness layers may propose learning candidates, but AMB owns policy evaluation, staging, suppression, and later review/promotion boundaries.

What Changed

• Added a v0 learning-candidate policy evaluator with explicit writeback decisions:
  • allow
  • needs_review
  • deny
  • degraded_no_write
• Added internal learning-candidate storage with candidate statuses:
  • pending
  • needs_review
  • approved
  • rejected
  • expired
• Recomputed learning policy at the storage boundary so callers cannot forge an allow decision.
• Stored candidate records as review-stage memory records tagged with kind:learning-candidate, candidate_status:*, authority_class:*, source_runtime:*, and schema tags.
• Hid learning candidates from ordinary recall, browse, export, and stats unless explicitly requested through review tags such as kind:learning-candidate or candidate_status:*.
• Documented the learning-candidate review queue in the authority contract.
• Kept the public MCP tool surface unchanged at 10 tools. There is no new learning_candidate, startup_packet, or task_packet MCP tool.

Evidence

Current release snapshot:

• pytest: 228 passed
• public MCP tools: 10
• deterministic proof: 4/4 checks
• learning-candidate tests cover:
  • safe candidate allow decisions
  • review routing for high-authority classes
  • denial for malformed, sensitive, raw-transcript, and unsupported candidates
  • degraded no-write behavior when AMB is unavailable
  • storage rejection for forged policy decisions
  • suppression from normal recall, browse, export, and stats
• signal contention benchmark:
  • signal_contention_case_count = 5
  • signal_contention_case_pass_rate = 1.0
  • unique_active_claim_rate = 1.0
  • duplicate_active_claim_count = 0
  • active_reclaim_block_rate = 1.0
  • stale_ack_blocked_rate = 1.0
  • stale_reclaim_success_rate = 1.0
  • pending_under_pressure_claim_rate = 1.0
  • initial_hard_expiry_cap_rate = 1.0
• adversarial memory-governance benchmark:
  • adversarial_case_count = 6
  • adversarial_task_count = 7
  • adversarial_governed_task_pass_rate = 1.0
  • adversarial_governed_blocked_record_leak_rate = 0.0

Boundaries

The public MCP surface remains the same 10 tools:

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

Learning candidates are staging records, not approved durable memory. A candidate can help a reviewer decide what to promote later, but it is not source-of-truth memory until a reviewed promotion or replacement path creates the final record.

This release does not add:

• a scheduler
• a watcher runtime
• a hosted backend
• an autonomous task runner
• a graph database
• a distributed queue
• a complete candidate review UI
• automatic durable writeback from raw transcripts

Next

Use 0.14.x for narrow release-polish work: docs alignment, small compatibility fixes, and review/promote ergonomics that do not expand the public MCP surface.

A future minor release should be reserved for a larger thesis such as pre-compaction capture, broader retrieval fixtures, or a complete reviewed candidate-promotion workflow.

v0.13.1 - Launch Surface Polish

v0.13.1 - Launch Surface Polish

Agent Memory Bridge 0.13.1 is a small polish checkpoint over the 0.13
coordination runtime.

Thesis

0.13.1 = stable launch surface for persistent engineering memory.

This release does not add new MCP tools or runtime behavior. It packages the
current main branch as a clean public checkpoint after documentation, security,
comparison, onboarding, and release-evidence hardening.

What Changed

• README and Chinese README now present the product as local-first MCP memory
  plus context compiler, not a Codex-only helper.
• Install examples are platform-neutral and keep OS-specific details out of the
  primary path.
• The before/after gotcha demo is easier to find from the README.
• Security, comparison, authority-contract, integration, and onboarding docs are
  aligned with the current public story.
• The release contract now checks test-count drift across the README,
  production status, and release announcement evidence.
• Tracked benchmark snapshot files are no longer hidden by .gitignore.

Evidence

Current release snapshot:

• pytest: 201 passed
• public MCP tools: 10
• deterministic proof: 4/4 checks
• signal contention benchmark:
  • signal_contention_case_count = 5
  • signal_contention_case_pass_rate = 1.0
  • unique_active_claim_rate = 1.0
  • duplicate_active_claim_count = 0
  • active_reclaim_block_rate = 1.0
  • stale_ack_blocked_rate = 1.0
  • stale_reclaim_success_rate = 1.0
  • pending_under_pressure_claim_rate = 1.0
  • initial_hard_expiry_cap_rate = 1.0
• adversarial memory-governance benchmark:
  • adversarial_case_count = 6
  • adversarial_task_count = 7
  • adversarial_governed_task_pass_rate = 1.0
  • adversarial_governed_blocked_record_leak_rate = 0.0

Boundaries

The public MCP surface remains the same 10 tools:

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

There are still no separate startup_packet or task_packet MCP tools. Startup
and task-time assembly remain context-compiler behavior behind the existing
surface while the packet shape continues to mature internally.

This release also does not add:

• scheduler
• watcher runtime
• hosted backend
• autonomous task runner
• graph database
• distributed queue

Next

Pause feature expansion and shift to distribution, dogfood, and user feedback.
The next engineering thesis should stay narrow: improve startup/task context
assembly without expanding the public MCP surface unless the internal packet
shape proves stable.

Post-release main update: harness boundary docs

After the v0.13.1 tag, main added a small documentation/spec checkpoint for the next GBrain-inspired direction:

• INSTALL_FOR_AGENTS.md for agent-readable setup
• llms.txt as a compact agent map of the repo
• docs/HARNESS-DESIGN.md to separate AMB Core from future harness tooling
• examples/harness-preview/ with sanitized startup/task packet preview shapes

This does not add new MCP tools or runtime behavior. The public surface remains the same 10 tools, and watcher/scheduler/harness execution stays outside AMB Core. CI for the follow-up commit passed across Ubuntu/macOS/Windows and Python 3.11/3.12/3.13.

v0.13.0 - Coordination Under Contention

v0.13.0 - Coordination Under Contention

Agent Memory Bridge 0.13 makes signal coordination easier to trust when several
agents or sessions poll the same namespace.

Thesis

0.13.0 = coordination under contention without turning the bridge into a task queue.

The public MCP surface remains the same 10 tools. The change is in the signal
contract: claim_signal(...) assigns work, extend_signal_lease(...) renews
ownership, and stale work must be reclaimed before it can be acknowledged.

What Shipped

• active same-owner claim_signal(...) calls no longer behave like lease renewal
• initial claims are capped by the signal hard expiry
• stale owners cannot ack_signal(...) after lease expiry
• generic claim selection filters eligible rows before the contention window, so
  active claims do not starve later pending work
• failed explicit claims return clearer reason codes such as already-claimed,
  claimed-by-other, expired, or lease-expired
• a signal contention benchmark slice for multi-consumer claim/reclaim/ack
  semantics
• release contract coverage for the new contention benchmark snapshot

Evidence

Current release snapshot:

• pytest: 194 passed
• public MCP tools: 10
• deterministic proof: 4/4 checks
• signal contention benchmark:
  • signal_contention_case_count = 5
  • signal_contention_case_pass_rate = 1.0
  • unique_active_claim_rate = 1.0
  • duplicate_active_claim_count = 0
  • active_reclaim_block_rate = 1.0
  • stale_ack_blocked_rate = 1.0
  • stale_reclaim_success_rate = 1.0
  • pending_under_pressure_claim_rate = 1.0
  • initial_hard_expiry_cap_rate = 1.0

Why It Matters

AMB is not only a memory store. It also carries lightweight handoff state. Under
repeated polling, stale leases, and competing consumers, that state needs to stay
legible. This release makes the ownership boundaries sharper without adding a
scheduler, worker runtime, or new MCP tools.

Honest Boundaries

0.13.0 does not mean:

• scheduler
• active worker runtime
• general queue platform
• distributed lock
• exactly-once distributed coordination
• dead-letter system

The contention benchmark is a repeatable serialized contract slice. It is not a
throughput benchmark or a claim that AMB is a full queueing system.

Next

Likely next candidates:

• pre-compaction capture before model-side context loss
• broader reviewed retrieval fixtures
• deeper real multi-client contention dogfood
• stronger write-side calibration for promotion quality

Pick one thesis for the next release instead of bundling them all together.

v0.12.2 - README Overview Correction

v0.12.2 - README Overview Correction

0.12.2 is a small hotfix on top of the 0.12 onboarding and integration
hardening work.

The runtime and public MCP surface stay the same:

• same 10 public MCP tools
• same platform-neutral setup and integrations docs
• same local doctor / verify onboarding checks
• same released validation snapshot with 185 passed

What changed:

• the README overview diagram now uses the intended first-screen artwork
• the public first impression is corrected without changing the runtime,
  benchmark claims, or MCP API

Why this hotfix matters:

• 0.12.1 improved the README surface, but not with the exact artwork that was
  meant to ship
• 0.12.2 fixes that mismatch directly
• this keeps the public surface honest and closer to the intended product story

v0.12.1 - README Overview Polish

v0.12.1 - README Overview Polish

0.12.1 is a small patch release on top of the 0.12 onboarding and
integration hardening work.

The runtime and public MCP surface stay the same:

• same 10 public MCP tools
• same platform-neutral setup and integrations docs
• same local doctor / verify onboarding checks
• same released validation snapshot with 185 passed

What changed:

• the README overview diagram was rebuilt so the first screen now matches the
  real product shape more directly
• the public narrative is clearer about the four layers:
  • clients
  • MCP surface
  • Agent Memory Bridge core
  • local proof layer

Why this patch matters:

• the old overview was accurate, but flatter
• the new overview makes the product easier to understand at a glance without
  inflating claims or widening the public API
• this helps adoption while keeping the release surface honest

v0.12.0 - Onboarding And Integration Hardening

v0.12.0 - Onboarding And Integration Hardening

Agent Memory Bridge 0.12 makes the first five minutes easier for people who are
not already inside the Codex dogfood loop.

Thesis

0.12.0 = onboarding and integration hardening without expanding the public MCP surface.

The bridge still exposes the same 10 MCP tools. The change is around adoption
confidence: platform-neutral install docs, client-specific config snippets,
generic stdio guidance, and local checks that prove the bridge works before it
is wired into a real project.

What Shipped

• platform-neutral setup across the public docs
• a client support matrix with careful status labels:
  • verified
  • documented
  • locally tested
  • supported
• generated config snippets for:
  • generic stdio MCP
  • Codex
  • Cursor
  • Cline
  • Claude Code
  • Claude Desktop
  • Antigravity
• agent-memory-bridge doctor for install diagnostics
• agent-memory-bridge verify for isolated stdio smoke validation
• onboarding contract checks for:
  • required docs
  • README linkage
  • generated config parsing
  • placeholder-safe public examples

Evidence

Current release snapshot:

• pytest: 185 passed
• public MCP tools: 10
• retrieval benchmark:
  • 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
• classifier calibration:
  • sample_count = 16
  • classifier_exact_match_rate = 0.875
  • fallback_exact_match_rate = 0.062
  • classifier_better_count = 13
  • fallback_better_count = 2
• onboarding contracts:
  • generated client configs parse cleanly
  • public onboarding docs are placeholder-safe
  • public surface and release-contract checks pass

Why It Matters

Good memory infrastructure does not help if the install path feels like private
maintainer folklore. This release makes the bridge easier to adopt as a general
MCP stdio server without turning it into a client-specific plugin project.

Honest Boundaries

0.12.0 does not mean:

• every MCP client is locally verified
• automatic client-config mutation
• IDE plugin development
• a new MCP tool surface
• remote deployment parity across every client

Next

The next useful pressure point is not more setup polish. It is stronger
coordination semantics under contention: retry boundaries, reclaim behavior, and
multi-agent handoff examples that stay small enough to fit the bridge's core.

v0.11.0 - Governed Procedure Memory

v0.11.0 - Governed Procedure Memory

Agent Memory Bridge 0.11 makes procedure memory safer to use at task time.

Thesis

0.11.0 = governed procedure memory without expanding the public MCP surface.

The public MCP surface remains the same 10 tools. The change is inside task
memory assembly: procedure records now carry lifecycle and boundary fields, and
governed packets can prefer trustworthy procedures while suppressing stale,
replaced, or unsafe ones.

What Shipped

• procedure governance parsing for:
  • procedure_status
  • when_not_to_use
  • prerequisites
  • failure_mode
  • rollback_path
• governed task-memory packets:
  • prefer validated procedures
  • keep draft procedures eligible but lower priority
  • keep legacy no-status procedures eligible with warnings
  • suppress stale, replaced, and unsafe procedures
• recall_first(...) now treats calm procedure, checklist, and runbook
  queries as local-first task-memory triggers
• a reviewed procedure-governance benchmark slice

Evidence

Current release snapshot:

• pytest: 175 passed
• public MCP tools: 10
• procedure-governance benchmark:
  • case_count = 7
  • flat_case_pass_rate = 0.429
  • governed_case_pass_rate = 1.0
  • flat_blocked_procedure_leak_rate = 1.0
  • governed_blocked_procedure_leak_rate = 0.0
  • governed_governance_field_completeness = 1.0

Why It Matters

Procedure memory is useful only if stale or unsafe playbooks do not keep leaking
back into task context. This release keeps procedures auditable in the store while
making the selected task packet more conservative.

Honest Boundaries

0.11.0 does not mean:

• procedure execution
• automatic procedure learning from raw transcripts
• a scheduler or worker runtime
• a new MCP tool surface
• a claim that broad retrieval ranking improved

Next

The next step is not another hidden automation layer. The next useful pressure
point is better dogfood observability: measuring how task packets, procedure
governance, and coordination signals behave across real sessions.

v0.10.0 - Relation-Aware Task Memory

v0.10.0 - Relation-Aware Task Memory

Agent Memory Bridge 0.10 makes task-time memory more connected and more
explainable without expanding the public MCP tool surface.

Thesis

0.10.0 = relation-aware task memory on top of the same small MCP bridge.

The public MCP surface remains the same 10 tools. The change is inside the
task-memory assembly path: procedures, concepts, beliefs, domain notes, and
supporting records can now be filtered and assembled with relation metadata in
mind.

What Shipped

• relation-aware task-memory assembly
• support-chain completion for task packets
• suppression of superseded, contradicted, expired, future, or invalid records
• explainable packet decisions for selected and suppressed records
• reviewed task-memory benchmark fixtures comparing flatter packets against the
  relation-aware path

Evidence

Current release snapshot:

• pytest: 159 passed
• public MCP tools: 10
• task-memory benchmark:
  • case_count = 5
  • flat_case_pass_rate = 0.2
  • relation_aware_case_pass_rate = 1.0
  • flat_blocked_item_leak_rate = 1.0
  • relation_aware_blocked_item_leak_rate = 0.0
  • flat_avg_packet_size = 3.8
  • relation_aware_avg_packet_size = 1.8

Why It Matters

This release makes memory more useful at the moment of work. Instead of returning
a flat set of locally relevant records, the bridge can now build a smaller packet
that keeps direct support, removes stale or unsafe support, and explains why
records were included or suppressed.

Honest Boundaries

0.10.0 does not mean:

• a graph database
• broad graph traversal
• a new traversal-oriented MCP API
• a claim that general retrieval ranking improved
• automatic procedure learning
• procedure execution

Next

The next step is governed procedure memory: making procedure records more
trustworthy with explicit lifecycle states such as draft, validated, stale,
replaced, and unsafe.

v0.9.0 - Applicable, Compositional Memory

v0.9.0 - Applicable, Compositional Memory

Agent Memory Bridge 0.9 keeps the same small MCP bridge while the engine underneath
gets more structured and more applicable.

Thesis

0.9.0 = applicable, compositional memory on top of a still-small MCP surface.

The public MCP surface is still the same 10 tools:

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

What changed is the engine behind that bridge.

What Shipped

• relation-lite metadata surfaced through recall, export, stats, proof, and health tooling
• benchmarked retrieval with precision@k, recall@k, MRR, and expected_top1_accuracy
• metadata-only telemetry plus client/model provenance
• first-pass task-time assembly over procedures, concept notes, beliefs, and linked supporting records
• the planned 0.8 structure and credibility work folded into the direct 0.9.0 release

Evidence

Current release snapshot:

• pytest: 152 passed
• deterministic proof: 4/4 checks passed
• canonical benchmark:
  • 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
  • relation_metadata_passed = true
  • duplicate_suppression_rate = 1.0

Reviewed classifier calibration snapshot:

• 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

Why It Matters

This release makes the bridge more useful without turning it into a bigger platform.

• Memory is more structured than transcript capture.
• Retrieval claims are benchmarked instead of guessed.
• Task-time recall can assemble procedures, concepts, beliefs, and linked support.
• The public integration contract stays small and inspectable.

Honest Boundaries

0.9.0 does not mean:

• a graph database
• full relation-aware traversal or ranking across the whole store
• automatic procedure learning from raw transcripts
• a scheduler or agent runtime
• cross-domain concept synthesis

Next

The next pressure points are broader reviewed fixtures, deeper cross-record assembly,
more deliberate procedure promotion, and pre-compaction capture before model-side loss.

v0.6.5 - Claim selection fairness

v0.6.5 Announcement

English

Agent Memory Bridge v0.6.5 is out.

This release tightens signal coordination by making generic claim selection fairer under contention.

What changed:

1. Claim selection is now fairer

• when signal_id is omitted, claim_signal(...) selects from the oldest eligible window
• it avoids immediately favoring the same consumer's stale claim when other pending work exists
• this reduces accidental poll-order wins without expanding the MCP surface

2. The fairness contract is now tested

• deterministic proof now checks the stale-reclaim fairness case
• signal lifecycle remains claim -> extend -> ack / expire / reclaim
• full suite remains green

3. Learning-quality work stays intact

• slice-aware classifier calibration remains in place
• benchmarked retrieval still holds on the canonical fixture

Current state:

• no MCP surface break
• fairer generic claim behavior
• 80 passed

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

Release:
https://github.com/zzhang82/Agent-Memory-Bridge/releases/tag/v0.6.5

简体中文

Agent Memory Bridge v0.6.5 发布了。

这次发布主要把 signal coordination 再收紧了一步：让泛化的 claim selection 在竞争场景下更公平。

主要变化：

1. claim selection 现在更公平了

• 当 signal_id 留空时，claim_signal(...) 会从最老的 eligible window 里选择
• 如果还有其他 pending work，同一个 consumer 不会优先把自己刚过期的 stale claim 又捡回去
• 这样能减少纯轮询顺序带来的偶然优势，而且不需要扩 MCP surface

2. 这条 fairness contract 现在有测试保护

• deterministic proof 现在会检查 stale-reclaim fairness case
• signal lifecycle 仍然是 claim -> extend -> ack / expire / reclaim
• 全量测试继续保持绿色

3. learning-quality 这条线保持不变

• slice-aware classifier calibration 继续保留
• canonical fixture 上的 benchmarked retrieval 也还在

当前状态：

• 没有 MCP surface break
• generic claim 行为更公平
• 80 passed

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

Release:
https://github.com/zzhang82/Agent-Memory-Bridge/releases/tag/v0.6.5