Open-source AI agent with memory

The AI agent that remembers, checks its own work and gets better with use. Version 1.2.0, Apache-2.0, on PostgreSQL 17.

Decision · toolingResolved Pin CI to Node 22 after the build failed on 20 Confidence0.72 Build failed on Node 20empirical Node 22 is the current LTS lineanalysis Same fix worked for Python 3.12pattern

npm ci && npm test node v22 · 142 tests passed

recall_deep4 results "CI node version" CI runs on Node 22fact Build failed on Node 20episode Upgrade Node in CIprocedure Pin CI to Node 22decision

Factsuperseded CI runs on Node 20 Replaced byCI runs on Node 22 Whya newer fact contradicts it Recallserves the newer one

Get Nous on GitHub

Every Nous decision on the record

Before it acts, Nous recalls similar past decisions and checks its guardrails. It records the choice with its reasons and confidence, then reviews it against what happened.

See how the Brain works

Nous / Decisions

TLtooling · 0.72Pin CI to Node 22 after the build failed on 20Reviewed2h
PRprocess · 0.64Retry the flaky upload test once before failingPending3h
ARarchitecture · 0.81Use cursor pagination for the export APIReviewed5h
SEsecurity · 0.55Ask before deleting the old staging bucketNeeds you6h
TLtooling · 0.77Cache pip wheels between CI runsReviewed1d
ARarchitecture · 0.68Split the weekly report into two subtasksPending1d
PRprocess · 0.70Summarise long build logs before sendingReviewed2d
SEsecurity · 0.83Rotate the API token kept in .envReviewed2d
PRprocess · 0.60Move the weekly digest to Monday morningPending3d
TLtooling · 0.74Lint Markdown files in the docs jobReviewed4d

toolingstakes: lowoutcome: successRecorded 2h ago

Pin CI to Node 22 after the build failed on 20

Confidence0.72

Similar decisions2 foundBoth resolved

Guardrails4 checkedNone blocked

Reasons

empiricalThe build failed on Node 20 with a native module error.
analysisNode 22 is the current long-term support line.
patternPinning Python to 3.12 fixed the same kind of drift.

Deliberation

recallFound the failed build episode and the Python pin.
intentChange .nvmrc and the CI image, then rerun.
outcomeThe build passed on Node 22; reviewed as a success.
calibrationConfidence and outcome feed the Brier score.

Six things Nous does. Each one is in the code you clone.

Remember

1. Keeps memory between sessions

Episodes, facts, procedures and censors stay in PostgreSQL and come back in later sessions.

CI node version

CI runs on Node 22fact

Build failed on Node 20episode

Upgrade Node in CIprocedure

Pin CI to Node 22decision

CI runs on Node 20Superseded by a newer factfact

Cache pip wheels between runsdecision

Release checklistprocedure

Docs job lints Markdownfact

Decide

2. Records decisions with confidence

Each significant choice is saved with its reasons and confidence, then reviewed against the outcome.

record_decision0.72

Retry the flaky upload test once

Reasonsempirical, patternpending review

Categoryprocessstakes: low

Similar decisionsRetry the deploy once, resolved

Guardrails4 checked, none blocked

OutcomeWaiting for the next three runs

Learn

3. Learns procedures from what worked

Good decisions, episode lessons and error recoveries become procedures once they pass a lesson gate.

Proceduregate passed

Upgrade Node in CIFrom 2 resolved decisions and 1 episode lesson

Evidence citedNot taught by an existing procedure

1. Read the failing job's log

2. Change .nvmrc and the CI image

3. Rerun the build on a branch

4. Open a pull request

Verify

4. Checks its claims against the ledger

Side-effecting tool calls go into a durable ledger. After each turn, what Nous says it did is checked against it.

Claim checkmatches

"The tests pass on Node 22"

bash · npm testexit 0

"I opened the pull request"matches

bash · gh pr createexit 0

"I emailed the summary"no record

CorrectionNo send_email call ran this turn

Ledger3 side-effecting calls this turn

write_file · notes/ci.mdlocal write

Act on its own

5. Works without being asked

Heartbeat checks, schedules, background subtasks and multi-step plans run without a message from you.

BackgroundToday

Heartbeat: health checkok

Schedule: build log summaryrunning

Plan: release notes, step 3waits for you

Subtask: check broken linksdone

Heartbeat: email checkno findings

Plan: release notes, step 4queued

Sleep cycleAfter two idle hourslater

Daily digest of findingsscheduled

Guardrails

6. Blocks risky actions in code

Censors and guardrails stop a matching action before it runs, so a rule holds when the prompt forgets it.

Censorblocked

No force-push to main

git push --force origin main

Steered toA branch and a pull request

CreatedAfter a failure the agent found

Guardrailno-high-stakes-low-confidenceCEL rule

Guardrailrequire-reasonsCEL rule

CensorAsk before deleting a bucketactive

CensorNo secrets in commit messagesactive

Nous beside a typical agent

Most agents are good for one conversation and blank the next. Nous is built to run for months.

A typical agent · new sessionno history

You"The build is failing again."

Agent"I have no earlier context. Can you paste the log?"

Session endsThe context is clearedgone

Nous · the same messageon the record

RecallThe last failed build and the Node 22 pinepisode

Reply"Node 20 broke a native module last time. Checking the CI image."

RecordThe decision, with reasons and confidencedecision

ReviewChecked later against what happenedlater

AreaMost agents and harnessesNous

LifetimeOne session, then the context resetsAlways on, with months of history in one database

MemoryA chat log, a notes file or a vector storeTyped memories linked in a knowledge graph, consolidated every night

DecisionsMade and forgottenRecorded with confidence and reasons, reviewed against outcomes

Saying "done"Taken on trustChecked against a ledger of the tool calls that really ran

RulesPrompt instructions it may ignoreCensors and guardrails that block the action in code

InitiativeWaits for your messageHeartbeat checks, schedules, background tasks and plans

ImprovementOnly when you edit its promptLearns procedures from what worked, each through a gate

GradingGrades itselfBackground work scored by a separate, deterministic scorer

The Nous loop in every turn

Nous frames the situation, rebuilds the context it needs, checks guardrails and past decisions, acts, compares what it claimed with what ran, then learns from the result.

A sleep cycle feeds what was learned back into recall.

See the architecture

Turn 14 · work viewclaims match

Sense"Why is last night's build failing?"

Framedebug, for a failing builddebug

Recall1 episode, 2 facts, 1 procedure, within budget4 items

Deliberate2 similar decisions, guardrails passed0.72

Actbash · npm ci, bash · npm test2 calls

Monitor"The tests pass on Node 22" matches the ledgerverified

LearnOutcome recorded, one fact updatedfact

SleepTonight: link the episode, resolve the old factlater

The agent runs on your machine

Clone the repository, put your Anthropic API key in .env and start the containers. Migrations run on startup.

Read the quickstart

$ git clone https://github.com/tfatykhov/nous.git
$ cd nous
$ cp .env.example .env  # set ANTHROPIC_API_KEY
$ docker compose up -d
✔ Container nous-postgres   Started
✔ Container nous-agent      Started
✔ Container nous-scorer     Started
✔ Container nous-telegram   Started
$ curl http://localhost:8000/health
{"status":"healthy"}
$ curl -X POST http://localhost:8000/chat \
    -H "Content-Type: application/json" \
    -d '{"message": "Hello, who are you?"}'