Munin reads every past Claude Code and Codex session on your machine, learns what works for you, and produces next-step suggestions tied to your strategy.
Private beta signup
Tested in the agents we actually use
In a normal session, the agent can burn a 1M-token window rereading stale detours just to recover your goals, current task, and constraints. Often 95% of that context is irrelevant. Munin Brain puts the live ask, project state, and watchouts into one compact surface first.
$ munin brain --format prompt
<session_brain provider="codex" built_at="2026-04-15T02:48:55Z">
<agenda>
- current ask: remove unsupported agent claims + show Munin Brain
</agenda>
<state>
- verified: beta signup flow deployed to muninmemory.com
- finding: live build source is .firecrawl/eag/builds/munin.sitesorted.co
</state>
<project>
- active ask: tighten product claims to Claude Code + Codex only
- codebase map: docs.html · features.html · index.html · pricing.html · resources.html · blog
</project>
<user_operating_model>
- profile: execution-heavy workflow; compact state beats transcript replay
- friction: front-load active ask + project capsule before wide search
</user_operating_model>
<guidance>
- start from the current ask
- avoid raw transcript dumps and stale cross-project noise
</guidance>
</session_brain>$ munin nudge
<next_nudge>
priority: P0
scope: billing-retry
action: Land retry-window fix before Thursday freeze
evidence:
- 4 past sessions · retry window debate
- 2 corrections · "window must match refund SLA"
- 1 open ticket #482
confidence: 0.82
expected_effect: unblocks 3 downstream tickets
first_verify: pnpm test -- src/billing/retry.spec.ts
</next_nudge>$ munin recall "refund SLA"
3 matches across 6 months:
▸ 2026-03-18 · feature/refund-window
"refund SLA = 72h Pro, 24h Standard"
session: claude-code · confirmed 2×
▸ 2026-02-04 · docs/billing
"see policy.md §3 for SLA table"
session: codex-cli
▸ 2026-01-22 · fix/retry-window
"retry window MUST match refund SLA"
session: claude-code · promoted to rule 0.96$ munin friction --agent codex --last 30d
142 sessions analysed · 3 patterns detected
rewrites_bun_to_npm 7× prevent?
reorders_imports 5× prevent?
misses_idempotency_lock 3× promote to rule?
$ munin promote "use bun, not npm"
✓ rule active for codex
✓ precedent: 7 corrections
✓ confidence: 0.91Most AI memory products store notes your agent may or may not read. Munin reads your whole session corpus, learns what works for you, and produces ranked actions backed by your evidence.
Munin reads every prior Claude Code and Codex session across all projects, all time. It journals, deduplicates, and reconciles every command, correction, redirect, and verification into one shared working history.
Munin builds the memory from what already happened.
Import your strategy document: goals, KPIs, initiatives, constraints. Munin compiles it into a deterministic kernel and joins it with your evidence corpus. Each nudge carries the task, why it matters now, supporting evidence, confidence, and expected effect.
The agent stops asking and starts proposing.
Action candidates earn their way to rules. Munin ranks by precedent count, success count, and failure count across every session. Candidates promote to assertions, assertions promote to rules, observe-only by default. Stale memory ages out. Re-confirmed memory strengthens.
Freshness, stability, and confidence scores on every claim.
Messy repo, messy agent. Munin maintains a per-project kernel: open loops, live claims, checkpoints, current recommendation, active risks. Every session re-entry arrives with a first question to answer and a first verification to run.
The agent hits the right target because the target is explicit.
Candidates become assertions, assertions become rules. Nothing promotes without a precedent count, a freshness score, and a confidence number.
Every live claim carries freshness, stability, and confidence. Stale memory ages out. Re-confirmed memory strengthens. You never wonder if a rule is still true.
Every promoted claim generates at least one of five action types.
You've been pushing on billing retry reliability for two weeks. Thursday is the merge freeze. Strategy says Q2 priority is "reduce failed-checkout spillover."
Land the retry-window fix before Thursday. It's the last open loop tied to your highest-priority metric. Evidence: 4 sessions, 2 corrections, 1 ticket. Confidence 0.82.
Three times across two months, the agent invented a Stripe webhook signature header that doesn't exist. Each time you corrected it by reading the real payload.
Add a pre-edit hook that blocks any Stripe webhook handler write until stripe listen payload is cited in the diff. Silent on everything else.
Every time you ship a migration you run the same five verification steps. You've redirected the agent through them twelve times.
Promote "pre-migration verification" to a skill. Five steps, ordered, with the exact commands. Invokable by name; auto-suggested when a migration file is staged.
Seven sessions agree: in this repo, integration tests must hit a real Postgres, not a mock. You've corrected mock-first approaches every time.
Promote to a rule: "integration tests use real Postgres." Attached evidence, freshness 0.94. Loaded into context on any test-file edit.
Codex keeps rewriting your TypeScript in the house style it prefers. Claude already follows your preferences. Two agents, one repo, different outputs.
Tell Codex specifically: "follow repo prettier config, do not restructure imports." Claude stays clean because it does not need the extra directive.
Resume and handoff briefs stay active while the latest replay/eval against a held-out corpus verifies. If proof drops, Munin falls back to a deterministic packet path.
Token reductions are a side effect, not the headline. The compressed-read surface exists so the agent can read more and decide better. Not so you save dollars on a model bill.
Resume briefs stay active while the latest replay against a held-out corpus verifies. If proof drops, Munin falls back to a deterministic path.
[13:42:04] Replaying 12 live claims against held-out corpus... [13:42:06] Reconciled 847 prior sessions across 2 active agents. [13:42:08] Freshness drift: <0.02 Stability drift: <0.01 [13:42:09] Proof packet: ✓ 12/12 claims verified [13:42:09] Signed: 2026-04-14T13:42:09Z resume brief accepted — all continuity checks passed.
Two tested surfaces
Friction reports identify repeated corrections, redirect statistics, and behaviour-change recommendations per agent. Munin keeps shared context in one place while still letting Claude and Codex receive different guidance when they need it.
Full continuity brief on session start. Prevention hooks for things you've corrected more than twice. Skills promoted from repeated workflows.
House-style directives where Codex diverges from your repo. Silence where it doesn't. Hooks re-expressed in the format Codex honours.
Munin Brain exposes the active ask, project capsule, user operating model, and guidance hints before the agent starts rereading the wider corpus.
Open a file and Munin loads only the rules that apply to it. Claude and Codex both read from the same kernel, while Codex-specific directives stay Codex-specific.
// retry window must match refund SLA (rule · 0.96 conf) export async function scheduleRetry(evt: RefundEvent) { const window = retryWindowFor(evt.reason); await idempotencyLock(evt.id); return queue.push(evt, { delay: window }); }
Munin is not publicly available yet. Leave your email and we’ll send early access invites, rollout notes, and the first stable build when it’s ready.
No. It's a local Rust binary that reads your sessions on disk and writes memory to your machine.
No. There's no outbound network call. Local-only, no account, no telemetry.
Claude's built-in memory is notes you write. Munin builds the memory itself from your session corpus and produces ranked actions tied to your strategy.
Every promoted claim carries freshness, stability, and confidence. Replay/eval proof-gates the cutover. If evidence drops, the claim falls back.
You import a strategy document. Munin joins it with your evidence and produces bounded nudges with evidence, confidence, and expected effect.
Yes. That is the point: one memory surface across the Claude Code and Codex sessions you're already running.
The last good brief is in .munin/journal as plain text. Open it, paste it into your agent, keep going. Munin is a local Rust binary. No server, no recovery flow.
Join the beta list and we’ll email you when Munin is ready for early access.