CogniKernel

Persistent, structured project memory for Claude Code and Codex (yet...). Your agent
stops re-deciding what you already decided. No extra LLM in the loop, no API keys, no
tokens billed, extraction runs on two small local encoder models, and nothing
leaves your machine.

Quickstart

Install from PyPI:

pip install "cognikernel[embedding]"   # recommended: + dense retrieval & the fine-tuned encoder runtime
# or: pip install cognikernel          # core (lexical-only)

Prefer an isolated global CLI? pipx install "cognikernel[embedding]" or uv tool install "cognikernel[embedding]".

Set up a project — run once from the project root:

cognikernel init .            # register the project + install hooks and MCP for Claude Code and Codex
cognikernel install-heads     # download the fine-tuned encoder models (~270 MB, one-time, sha256-verified)
cognikernel doctor .          # verify everything is wired up

Then just start a Claude Code session in the project. Decisions are captured
automatically when the session ends and injected as a compact memory block at the
next session start — nothing else to do. (install-heads is optional but
recommended: without it, extraction falls back to a weaker lexical path, and
doctor tells you which is active — see Setup details.)

Persistent, structured project memory for Claude Code — and Codex.
CogniKernel watches a coding session through its hook surfaces, extracts the decisions, constraints, and abandoned approaches worth keeping, consolidates them into an event-sourced store, and injects them back as a compact context block the next time you work — so the agent stops re-deciding what you already decided.

The store is keyed on the project path, so memory made in one tool travels to the other. It is not a vector-database wrapper. It is an event-sourced log of typed memory with lexical-primary retrieval, write-time consolidation, and a fail-open reliability spine designed never to break your session. And there is no extra LLM in the loop. Most memory tools work by sending your transcripts to a generative model to "summarize what mattered" — another API key, per-session token cost, added latency, and your session content leaving the machine.

CogniKernel treats extraction as classification, not generation: a deterministic sanitize → classify → consolidate pipeline, with two small fine-tuned encoder models (~130 MB ONNX, run locally on CPU in milliseconds)scoring salience and detecting when a new decision supersedes an old one. No API calls, no tokens billed, nothing leaves your machine. The only LLM involved is the coding agent you already run — CogniKernel makes it remember.

What you get, out of the box:

• Automatic capture & recall — decisions, constraints, and abandoned
  approaches are extracted at session end and injected next time. You never
  hand-write memory to CLAUDE.md.
• Cross-tool memory — one store serves Claude Code and Codex working in the
  same directory (Cross-platform).
• Action-time guardrails — a prior "don't do X" surfaces at the moment you're
  about to do X, not three files later (Claude Code).
• Read efficiency — an injected AST skeleton means the agent re-reads far
  fewer files (What it saves you).
• MCP tools for targeted use: recall · find_related · skeleton ·
  get_session_state.
• CLI: init · doctor · install-heads · codex-sync · show · reset
  (full list under Interfaces).
• No LLM, no cloud, no keys — everything runs locally; nothing leaves your
  machine.

The fine-tuned models

Two small encoder models do the actual thinking behind CogniKernel — not a keyword
list, not an LLM call:

Validated on held-out data, not vibes. Each model is scored against a frozen
evaluation set mined from real captured project stores that the training data never
sees — retraining on the eval set is a standing rule violation, not a shortcut we
occasionally allow. Because the underlying task (is this sentence a decision, or just
description?) is itself genuinely ambiguous at the margin, we also measure how much
two independent human labelers agree with each other on the same sentences, and hold
the model to that ceiling rather than to a naive 100%. The model's minority classes
(rare event types with few historical examples) are tracked and grown deliberately,
because a thin eval slice can hide the exact class that most needs work.

They're optional but on by default — here's the fail-open contract. cognikernel init
writes a per-project config that already selects the fine-tuned path
(extractor = "v2-broad", cross_encoder_supersession = true). What's missing after
init is the ~270 MB of model weights themselves — that's what
cognikernel install-heads fetches. Until you run it, CogniKernel works, but degrades
both decisions to a deterministic keyword/lexical fallback (weaker: it can miss
decisions phrased outside its trigger vocabulary, and it can't catch a paraphrased
correction lexical matching misses) — cognikernel doctor names this state explicitly
rather than leaving it invisible (see below). Nothing breaks either way — that's
the fail-open design — but the fine-tuned path is the better-quality one, and the
one new projects are configured to use by default. See Quickstart to
install them.

The memory loop

Everything CogniKernel does is one loop: observe → extract → consolidate →
store → retrieve → assemble → inject. The left rail captures; the right rail
recalls; the spine underneath keeps both honest.

   +========================= CLAUDE CODE SESSION =========================+
   | working memory  -  the context window the agent reasons in            |
   +--------------+-----------------------------------------+--------------+
                  | Stop hook captures transcript           | inject block
                  v                                         ^
   +--------------+---------------+         +---------------+--------------+
   | [1] EXTRACTION PIPELINE      |         | [6] COMPRESSION + INJECTION  |
   |   sanitize -> classify ->    |         |   authority-weighted budget, |
   |   salience (ONNX) ->         |         |   drop-to-fit (keep every    |
   |   decision-key + contracts   |         |   constraint) -> block       |
   +--------------+---------------+         +---------------+--------------+
                  | enqueue                                 | rank + fit
                  v                                         ^
   +--------------+---------------+         +---------------+--------------+
   | [2] WORKER + CONSOLIDATION   |         | [5] RETRIEVAL                |
   |   claim -> delta-merge ->    |         |   FTS5 BM25 + dense -> RRF   |
   |   supersede (latest-wins) -> |         |   prohibition_search (K1)    |
   |   project   (idempotent)     |         |   skeleton graph (PageRank)  |
   +--------------+---------------+         +---------------+--------------+
                  | persist (atomic)                        | recall
                  v                                         ^
   +--------------+-----------------------------------------+--------------+
   | [3] EVENT-SOURCED STORE   *   SQLite (WAL)                            |
   | typed events | evidence | provenance | FTS5 | embeddings | ledger     |
   +----------------------------------------------------------------------+
   | [4] RELIABILITY SPINE   atomic migrations | idempotent replay |       |
   | doctor --strict | fail-open hooks | import-linter | CI gate           |
   +----------------------------------------------------------------------+

The four hook surfaces

CogniKernel attaches to a session at four points. Each is fail-open — if memory
is unavailable or errors, the hook logs at WARNING, returns cleanly, and the
session continues. salience_v2 and supersession_xenc (above) do the
classification behind Capture and Session block; nothing here is a hardcoded
keyword list unless the fallback path is active.

What gets remembered

Memory is typed, not free-text chunks. The type drives ranking, rendering,
and supersession:

• DECISION — a choice that was made ("use Redis for the rate limiter")
• CONSTRAINT_HARD / CONSTRAINT_SOFT — rules, graded by deontic force
• APPROACH_ABANDONED_DO_NOT_RETRY — a dead end, kept in the graveyard so
  the agent doesn't re-attempt it
• conventions, config facts, schema decisions (canonical role keys)

A decision key lets a later restatement supersede an earlier one
(latest-wins), so the store self-consolidates instead of accumulating
contradictions. An optional cross-encoder adds semantic supersession when the
encoder backend is installed.

Retrieval

Lexical-primary, with dense as a fused signal — never pure vector:

• Hybrid core — FTS5 BM25 ∪ optional dense embeddings → Reciprocal Rank Fusion
• prohibition_search — a type-restricted lexical pool so a "do not do X" rule
  can't be crowded out by topically-similar prose at the moment you're about to do X
• Skeleton — an AST symbol graph ranked by PageRank; find_related unions
  semantic (embedding) neighbours with import-graph-adjacent events to
  surface what a change touches (the semantic axis needs the embedding extra)
• Golden-record consolidation at read — latest-wins reconciliation so recall
  returns one coherent answer, not a pile of revisions

What it saves you

Benchmarked in a three-arm comparison — CogniKernel vs flat curated notes vs no
memory — with real agent sessions across four multi-session projects. Full
methodology, per-project tables, and the honest caveats (including where
CogniKernel ties or loses) are in docs/benchmark.md:

• File reads: the universal win. The CogniKernel arm made the fewest file
  reads in every project — typically 2–4× fewer (23 vs 63, 16 vs 47/53,
  40 vs 89/83), and in the best case 3 reads vs 29 because the injected
  block + AST skeleton carried the whole repo's shape. Fewer reads means fewer
  tool round-trips and more of the context window left for actual work — your
  session gets longer before compaction, not just cheaper.
• Tokens: ~30–40% leaner than auto-memory where memory matters. On projects
  with evolving decisions and cross-file dependencies, CogniKernel used roughly
  30–40% fewer tokens than a native auto-memory arm — and stays ahead once
  price-weighted (cache-read is billed 0.1×, so ~95% of any session's bill is
  discounted cache-read; the weighted edge is real but smaller than the raw
  figure). Roughly a wash on small implementation-heavy projects. Precise
  per-project figures are pending a single-model re-benchmark — see
  docs/benchmark.md.
• Recall instead of re-derivation. Where memory earns its keep is projects
  whose state is too large, too evolving, or too long-lived to re-derive
  cheaply: the agent starts already knowing the decisions, constraints, and
  dead ends, instead of spending the first quarter of the session rediscovering
  them.

Reliability — the spine

The system is designed to degrade legibly, never silently:

• Atomic migrations — each numbered migration applies its body + version bump
  in one transaction; safe to crash mid-script
• Idempotent replay — a re-run worker job can't double-count or drift decay (evidence-provenance guard)
• Fail-open hooks — every surface swallows its own failure and logs at WARNING; silence never reads as success
• cognikernel doctor --strict — per-subsystem health (schema, FTS5, embeddings, symbols, worker queue); non-zero exit when degraded
• Architecture enforcement — import-linter layered contracts, guarded by a meta-test so a typo can't silently disable them
• CI promotion gate — lint + full suite (incl. tests/reliability/ failure-injection) on every PR; see CONTRIBUTING.md

Cross-platform (Codex)

The store is platform-neutral — one SQLite DB per logical project, so Claude Code
and Codex working in the same directory share one memory. Project resolution is
alias-aware: C:\repo and /mnt/c/repo resolve to the same store, so memory
follows the checkout across Windows, WSL, and native mounts; for genuinely
different checkout paths, an opt-in project_identity key in
.cognikernel/config.toml pins them to one shared store. Codex reads memory through
the registered MCP server (get_session_state / recall); the capture direction
is pull-based, because Codex has no Stop-hook equivalent:

• cognikernel codex-sync <project> scans ~/.codex/sessions for rollouts whose
  recorded cwd maps to the project and captures the delta through the same
  extraction pipeline (a rollout→transcript adapter is the only Codex-specific
  code; delta/dedup/idempotency are shared and unchanged).
• Automatic at the handoff — Claude's SessionStart drains pending Codex
  rollouts before building the block, and the MCP server's queue drainer pulls
  new rollouts each cycle, so a live Claude session picks up Codex-side decisions
  without waiting for the next session; on the Codex side, init writes an
  AGENTS.md instruction + a ck-sync skill so Codex pulls at session start.
• init provisions both — .mcp.json (Claude) and .codex/config.toml
  (with the server's cwd + project env pinned) + AGENTS.md (Codex),
  idempotently and without clobbering existing settings.
• cognikernel doctor reports a codex health line (sessions dir + rollout
  count, or "nothing to sync" — Codex is optional, so its absence is healthy).

A decision made in Codex reaches the next Claude session's block, and
vice versa. The action-point surfaces (CK-1, PreToolUse gate) are Claude-only —
Codex has no per-prompt/per-tool hook — so on Codex the loop degrades to the shared
block + MCP recall.

Interfaces

MCP tools (the session block is injected automatically; these are for targeted use):
recall · find_related · skeleton · get_session_state

CLI:

• cognikernel init <project> — register the project and install the session hooks
• cognikernel doctor [--strict] <project> — subsystem health report
• cognikernel codex-sync <project> — capture Codex CLI sessions for this project
• cognikernel install-heads — install the trained encoder artifacts (salience + cross-encoder ONNX bodies): downloaded from the heads-v1 release and sha256-verified, or copied from a local models/ export when present
• cognikernel show <project> / cognikernel reset <project> — inspect / clear stored memory

Setup details

The Quickstart above is the whole flow; this expands on the two
steps worth understanding.

install-heads is the easiest to skip and most worth not skipping.
cognikernel init already configured the project to use the fine-tuned path
(extractor = "v2-broad", cross_encoder_supersession = true in
.cognikernel/config.toml) — install-heads is what actually supplies the model
weights those settings call for. Skip it and CogniKernel still runs, just on the
weaker deterministic fallback (fail-open, not a hard error). The download pulls
from the heads-v1 release
and is sha256-verified; pass --source <dir> to install from a local models/
export instead, or --no-download to rely on the fallback deliberately. The
fine-tuned runtime (onnxruntime + tokenizers) ships with the [embedding] extra,
so install with pip install "cognikernel[embedding]" for the full path.

cognikernel doctor tells you which path is active — its subsystem-health
block reports salience_head and supersession_head by name:

-- subsystem health -----------------------------------------
  [OK] salience_head    : installed, loads from ~/.cognikernel/models/salience_v2
  [OK] supersession_head: installed, loads from ~/.cognikernel/models/supersession_xenc

If you skipped install-heads (or installed core-only, without the [embedding]
extra that provides the ONNX runtime), the same lines say not installed and
name the exact command to fix it — always informational, never a doctor --strict failure, since the legacy fallback is a fully supported mode.

Working from a clone instead of PyPI? Substitute uv sync --extra embedding
for the install and prefix the commands with uv run (e.g. uv run cognikernel init .).

Project layout

src/cognikernel/
  integration/   hooks, CLI, MCP server, session/worker orchestration
  extraction/    sanitize -> classify -> salience -> decision-key pipeline
  delta/         delta-merge + supersession (latest-wins; cross-encoder optional)
  retrieval/     hybrid BM25 + dense -> RRF
  storage/       event-sourced SQLite, FTS5, migrations, render ledger
  embedding/     optional dense vectors (fastembed)
  symbols/       AST skeleton + PageRank graph
  compression/   authority-weighted drop-to-fit budget
  injection/     block template assembly
  model.py       Event — the dependency-free domain primitive
tests/
  unit/          per-subsystem
  reliability/   crash-replay · worker-contention · corrupt-input injection

Status

Schema v18 (includes the Codex cross-platform capture). Architecture
contracts: 3 kept / 0 broken. CI gate: lint + full suite on Ubuntu (3.11/3.12) and
Windows. See CONTRIBUTING.md for the Definition of Done that gates every change.