CogniKernel

Your AI coding agent has amnesia by design. Every session starts from zero — it re-reads the same files, re-derives the same architecture, and re-asks questions you already answered last week. CogniKernel gives the project a memory that outlives any one session, so the agent stops paying the same cost twice:

  WITHOUT MEMORY                          WITH COGNIKERNEL
  ───────────────                          ─────────────────
  Session 1                                Session 1
    "explain the auth flow"                  "explain the auth flow"
    -> reads 12 files, learns it             -> reads 12 files, learns it
    -> session ends, all of it lost          -> session ends
                                              -> decisions captured & stored

  Session 2                                Session 2
    "explain the auth flow"                  already knows: the auth flow,
    -> reads 12 files again                  why Redis beat an in-process
    -> re-derives the same thing             cache, and the approach that
    -> session ends, nothing kept            was tried and abandoned
                                              -> picks up where it left off

  Session 3                                Session 3
    "explain the auth flow"                  already knows everything
    -> reads 12 files a third time           sessions 1 and 2 settled
    -> context resets again                  -> starts from where YOU are,
                                                 not from zero

No extra LLM in the loop, no API keys, no tokens billed for memory itself, 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     # optional, recommended: 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.)

---

What's new in 0.1.3

• Fresh installs get the memory tools again. 0.1.2 accepted mcp 2.x, which removed a module the MCP server needs, so recall, find_related, skeleton and get_session_state could be missing. 0.1.3 requires mcp<2. On 0.1.2, run pip install "mcp<2".
• The right open work item. The block's "Working on" line is no longer taken over by Claude's step narration, an ordinary instruction, or a sentence that only points at a work item, and it keeps its place when decisions pile up. A work item carried over from an earlier session says which one: (S1 · 09-12).
• Less of Claude's own chatter kept as memory. Step narration ("Now let's run the full test suite."), answer lines ("Summary: …") and narration about CogniKernel itself now rank below real decisions or are no longer stored.
• Memory follows you into subdirectories of the same git repository instead of starting a second, empty store.
• Ask memory why:
  • cognikernel why: where a claim came from, why it ranks where it does, and what it replaced
  • cognikernel show --as-of: what memory believed at a past date or commit
  • cognikernel explain-recall: why recall found what it did
• Cheaper by default: new projects start with the advisory read gate. Under the old strict default, 89% of refused first reads were retried at once. Existing projects keep their setting.
• Corrected numbers: cognikernel telemetry, doctor and the published benchmark counted token usage once per transcript line instead of once per response.

Details: CHANGELOG.md · release notes.

---

What CogniKernel actually is

AI coding agents don't remember anything between sessions by default — the model's weights are frozen, so its only memory is whatever fits in that session's context window. Two consequences of that are common enough now to have names: context rot, where a long session's own window fills with stale tool output and superseded instructions until quality degrades, and cold-start re-derivation, where the moment a session ends, all of that understanding is gone and the next one pays the "getting up to speed" cost again. At scale, that repeated cost is a real reason AI coding spend has become something engineering orgs actively budget and worry about — this project is one attempt at closing that gap.

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. 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, and it is not another LLM summarizing your transcripts. Most memory tools work that way — send your session to a generative model to "summarize what mattered," which means 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, milliseconds on CPU) 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 · why · explain-recall · 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:

Model	What it decides	Why it's an encoder, not an LLM
salience_v2	For every sentence in a session transcript: is this worth keeping, and what type — DECISION, CONSTRAINT_HARD/SOFT, APPROACH_ABANDONED_DO_NOT_RETRY, or noise? This is the model that decides what gets remembered.	A fine-tuned SetFit classifier (bge-small backbone, ~130 MB ONNX) runs this on CPU in milliseconds. An LLM doing the same job costs an API call and real latency on every session — CogniKernel spends neither.
supersession_xenc	Does this new fact replace an existing one (a decision changed, a constraint was relaxed), or does it just restate what's already stored? This is the model behind latest-wins consolidation.	Same constraint: this fires on every capture. A cross-encoder scores it locally in milliseconds instead of round-tripping to a model API.

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.

  SessionStart          UserPromptSubmit         PreToolUse            Stop
  ─────────────         ─────────────────        ──────────           ──────
  inject the block  ──▶  surface relevant    ──▶  warn before a   ──▶  extract &
  memory picks up        memory as you type       past decision        persist what
  where you left off     it, unasked               gets violated        this session
                                                                         decided

Surface	Hook	Authority	What it does	Why you care
Session block	SessionStart	advisory	injects the canonical decisions/constraints/skeleton block	a new session already knows what the last one decided — no "let me re-read the codebase to remember where we were"
CK-1 recall	UserPromptSubmit	advisory	surfaces prompt-relevant memory, dual-evidence gated, dedup'd via render ledger	ask about a subsystem and the relevant prior decision rides in with your prompt, unasked
Read/Edit gate	PreToolUse	advisory (strict opt-in) / JIT	read efficiency on Read/Grep: refuses re-reading a file already read this session, and refuses first reads only under hook_policy = "strict"; just-in-time prohibition surfacing on Write/Edit (K2)	the agent gets warned at the moment it's about to violate a past decision, not three files later when you notice
Capture	Stop	side-effect	extracts and persists decisions — you never write memory to CLAUDE.md by hand	you never write down what you decided; the next session already has it

---

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:

  Orientation reads — reads before the first line of code, i.e. the cost of
  working out where you are. CogniKernel vs. Claude Code native auto-memory.
  (bar scale: 20 chars = 40%; see docs/benchmark.md §3)

  project stresses…      CK / auto    reduction vs auto-memory

  evolving decisions     27 / 39   ███████████████░░░░░  -30.8%
  small, re-readable     17 / 20   ████████░░░░░░░░░░░░  -15.0%
  quality invariants     97 / 104  ███░░░░░░░░░░░░░░░░░   -6.7%
  self-authored API      90 / 92   █░░░░░░░░░░░░░░░░░░░   -2.2%

• Orientation reads: the one win that holds everywhere. Lower on 4 of 4 projects. This is precisely what an injected block is for — the agent starts already knowing the repo's shape instead of paying to rediscover it. Fewer round-trips also means more of the context window left for actual work, so your session runs longer before compaction, not just cheaper.

• Total reads and tokens: a split result, not a sweep. Once you count every read rather than just orientation, CogniKernel is ahead on one project of four (Relay, −38.1%) and slightly behind on the rest. Against native auto-memory, price-weighted cost is −25.0% on Relay and −19.3% on Toolbelt, and within a few percent on Conductor (+3.0%) and Taskflow (+4.2%). Against no memory at all it is not a sweep: on Relay CogniKernel costs 23% more than the no-memory arm, from round-trips its own tools add. Structured memory repays its overhead on projects with several evolving decisions and several abandoned approaches; on small re-readable ones it adds little. (Corrected 2026-09-10 — earlier figures double-counted token usage; see docs/benchmark.md §3.)

• 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> [--as-of <date|commit>] / cognikernel reset <project> — inspect stored memory (or what it believed at a past date or commit) / clear it
• cognikernel why <project> <#id or words> — where a claim came from, why it ranks where it does, and what it replaced or was replaced by
• cognikernel explain-recall <project> <query> [--claim #id] — why recall and the per-prompt push retrieved what they did, or didn't
• cognikernel telemetry <project> — count Claude Code session usage (once per API response) for doctor's cache and round-trip figures

---

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
  quality/       admission gate + defect detectors (a dependency-free leaf)
  telemetry/     Claude Code usage ingest: cache and round-trip figures
  utils/         shared helpers
  model.py       Event — the dependency-free domain primitive
tests/
  unit/          per-subsystem
  reliability/   crash-replay · worker-contention · corrupt-input injection

---

Status

Version 0.1.3. Schema v22 (adds per-response telemetry and round-trip counts, on top of the supersession/archival timestamps and commit anchor that show --as-of and why read). Architecture contracts: 4 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.