Brain-AI Memory v0.6.0 — Automatic Session Memory for Codex and Claude Code

Brain-AI Memory can now keep a project moving across Codex and Claude Code sessions without requiring you to rebuild the context by hand each time.

What’s new

• Opt-in, project-scoped automatic session memory for Codex and Claude Code
• Bounded recall at session start and before prompts
• Approved procedural-rule checks before actions
• Successful-edit receipts and dirty-only checkpoints
• A next-session handoff that can resume unfinished work
• Preview-first install, exact managed-hook ownership, and safe disconnect

Try it

Install the tagged release:

git clone --branch v0.6.0 --depth 1 https://github.com/Hahyun-Lee/brain-ai-memory.git
cd brain-ai-memory
python3 -m venv .venv
source .venv/bin/activate
python -m pip install ".[mcp]"

Then, in the project you want the agent to remember:

cd /path/to/your/project
export PROJECT_ROOT="$PWD"
export BRAIN_AI_HOME="$PROJECT_ROOT/.brain-ai"

brain-ai init
brain-ai entity add --name my-project --type project

# Preview the project-scoped Codex integration, then apply it.
brain-ai connect codex --entity my-project --mode loop --project-root "$PROJECT_ROOT"
brain-ai connect codex --entity my-project --mode loop --project-root "$PROJECT_ROOT" --apply
brain-ai doctor --host codex --entity my-project --mode loop --project-root "$PROJECT_ROOT"

Use claude-code instead of codex for Claude Code. The first connect command is a preview; the second writes the project configuration.

Evidence boundary

The v0.6.0 suite contains 123 tests covering the loop, host integration, storage durability, concurrency, restart/resume, failure injection, and clean-wheel execution. These checks establish packaging and deterministic integration behavior. They do not claim improved LLM answer quality or long-run reliability in every host environment.

See the v0.6.0 changelog and automatic session memory guide.

Brain-AI Memory v0.5.0 — Project Memory That Survives the Next Session

Project memory that survives the next session

Brain-AI Memory v0.5.0 turns an existing MEMORY.md into reviewed,
project-scoped memory that an MCP host can recall, update, checkpoint, and
resume. The source file stays untouched, replaced facts keep their history, and
exact state no longer has to be guessed from prose.

What is new

• Audit a Markdown memory file before anything enters the typed store.
• Approve, apply, and logically roll back imports with file-and-line provenance.
• Keep project-bound memory, rules, state, and handoffs out of other projects.
• Write preview-first project MCP configuration for Codex or Claude Code.
• Resume the latest handoff after a fresh MCP process starts.
• Inspect local operation without an API key, hosted service, or database server.

Verified release path

• 64 public software tests.
• Python 3.10, 3.11, and 3.12 compatibility checks.
• Clean-wheel audit → review → apply → generated Codex MCP configuration → real
  stdio tool calls → checkpoint → fresh process → resume.
• Runtime contract: 14/14; component contracts: all-ten 20/20 versus flat 1/20.

These are software and integration results. They do not claim that Codex or
Claude Code will always choose the tools, or that this release improves LLM
answer quality over RAG or long context.

Try it

git clone --branch v0.5.0 --depth 1 https://github.com/Hahyun-Lee/brain-ai-memory.git
cd brain-ai-memory
python3 -m venv .venv
source .venv/bin/activate
python -m pip install ".[mcp]"
brain-ai --home "$(mktemp -d)" tour

Then follow the adoption workflow
or the 한국어 README.

Use it when work spans many sessions, facts or state keep changing, and a stale
or cross-project memory can cause a real mistake. For a short one-off task, a
plain MEMORY.md is still the simpler choice.

Brain-AI Memory v0.4.0

Brain-AI Memory v0.4.0 centers the public package on its implemented core: an installable, local, provider-neutral reference kernel for typed operational memory in long-running agents.

Highlights

• Ontology v3 requires explicit memory or control categories, validates category values, and reports observable counts; the bundled schema declares the canonical five-memory plus two-supporting-control split.
• Optional entity scope now reaches harness and fallback sequence gates, so supplied entity-bound rules are evaluated at the execution boundary.
• Ablation reproducibility now verifies normalized semantic outcomes and recorded-source provenance for the ten tested mechanisms.
• English and Korean documentation, the graphical abstract, and the social preview now separate the managed-memory path from the optional memory-to-action bridge.
• Runtime, package metadata, and citation versions now share a CI-verified release contract.

Compatibility

Custom ontology v2 files must add category: memory or category: control to every component before upgrading.

Evidence boundary

The 20/20 versus 1/20 ablation result is authored contract conformance, not evidence of better LLM answer quality or superiority over RAG. Automatic transcript ingestion, token-budgeted context injection, autonomous lifecycle scheduling, physical deletion, and production action enforcement remain host responsibilities.

Download the wheel below and install it with:

python -m pip install ./brain_ai_memory-0.4.0-py3-none-any.whl

Brain-AI Memory v0.3.1

Patch release for Smart Connections users. The semantic adapter now keeps one stdio server process alive across queries, so embedding and reranker models stay warm; a live two-query smoke reused the same process and reduced total latency from 5.9 s cold to 1.4 s warm. Adds explicit mcp_env configuration for SMART_SEARCH_PROFILE=adaptive, retains v1/v2 response compatibility, preserves v2 server ranking, and closes failed child connections before local fallback. CI, 16 tests, MCP round-trip, runtime contract 14/14, and 420-record component ablation passed.

Brain-AI Memory v0.3.0

Installable provider-neutral MCP control layer with scoped memory, entity/relation bindings, exact state, deterministic action gates, lifecycle handoffs, and a one-minute tour. Adds v1/v2 Smart Connections adapter compatibility, bounded v2 snippets, and server-ranked hybrid retrieval without duplicate BM25. Includes the 20-case, 21-condition component-contract ablation (420 records). Evidence boundary: this validates public software contracts, not biological causality or end-to-end LLM QA superiority.

Brain-AI Memory v0.2.0 — Installable Public Runtime

Brain-AI Memory is no longer an architecture-only extraction. v0.2.0 ships an installable, local-first public alpha.

What is now executable

• differentiated HC episodic, ATL semantic, BG rule, IPS numerical, TH gate, CB harness, and PFC routing paths
• checkpoint, consolidation preview/apply, reconsolidation, and seven lifecycle operations
• deterministic fallback sequences that continue until success, block, or exhaustion
• local multilingual BM25, Markdown vault, and Smart Connections MCP adapters with reciprocal-rank fusion and visible fallback diagnostics
• localhost-only clean-room Command Center and read-only JSON APIs
• English and Korean runtime documentation

Install

python3 -m venv .venv
source .venv/bin/activate
python -m pip install brain_ai_memory-0.2.0-py3-none-any.whl
brain-ai init
brain-ai demo
brain-ai status

Verification

• 8 runtime/adapter tests pass on CI
• 14/14 runtime component-contract cases versus 8/14 for a flat retrieval-only control
• clean-environment wheel install and end-to-end fallback execution verified
• actual Smart Connections MCP compatibility verified with MCP + vault fallback fusion

The 14-case A/B is contract conformance, not evidence of improved LLM QA or general agent performance. Release-grade LongMemEval/MemoryAgentBench QA evaluation remains pending. The observer has no authentication and should remain bound to localhost.

Brain-AI Memory v0.1.0

First public release of Brain-AI Memory.

• Diagnostic architecture for separating memory, routing, rule, execution, numerical-state, and input-gating failures
• Runnable standard-library examples and reusable lifecycle, rule, and hook templates
• Sanitized operational evidence with explicit causal limits
• Reproducible 500-question LongMemEval-S retrieval pilot, including negative results and auditable artifacts
• CI, security reporting, dataset attribution, and citation metadata

The public repository is a clean-room extraction. It does not include the private operational backend or raw operational records.