Give your AI a brain
that persists.
Null Memory makes every session feel like talking to the same person. Facts, decisions, mistakes, reflections — your agent remembers and grows.
Aleph gives it eyes. Null gives it a brain.
New
Build your AI organization
One agent that remembers was the start. Null is now the toolkit for building your own AI organization: multiple seats — each a personality with its own memory store and its own growing identity — exchanging knowledge over typed edges. You define the structure, name the personalities, and set the authority levels. Null ships the primitives.
Built and battle-tested running our own multi-seat org.
Personalities, not clones
null persona create builds a new seat's identity through a question-driven interview — replayable with --answers — and persona onboard brings it into the org. From there each personality accumulates its own anchors, voice, and working relationship with its human.
Identity tiers
A trust ladder from anonymous install to full identity. Promotion is an explicit decision; offboarding is revoking access. A seat's working set was only ever what its tier allowed — by construction, not by cleanup.
The exchange
Your org's hallway: a shared git repo of per-seat append-only streams carrying reports, broadcasts, advisory claims, and questions. Each seat writes only to its own stream, and nothing private enters the hallway — privacy by construction, not by filter.
Reports, not keystrokes
Workers don't stream observations upward — they ship consolidated session reports. Every ingested report carries provenance and a confidence discount, so another seat's claim is never confused with first-hand knowledge.
Onboarding packets
New seats get a curated packet — decisions-in-force and project context scoped to their role, with identity content excluded by default — not a replay of company history.
Decision authority
Decisions carry an authority level — board, org, team, seat are the defaults; the names and topology are your configuration. Contradictions between seats are detected with provenance and resolved by an explicit decision that flows back down.
Multi-machine sync. Conflict-free by construction.
The org rides on event-sourced sync: announcements travel between seats, artifacts stay in their homes, and knowledge stays in each seat's own store.
Event-sourced sync
Every writer appends only to its own log; the database is a local view replayed from the union. Merge conflicts are structurally impossible, and git is the default transport — free, private, offline-first, fully auditable.
The UDP doorbell
After every push, a contentless ping tells peers to fetch now — seconds-latency delivery on a LAN. The ping carries and trusts nothing; the periodic poll is the delivery guarantee, the ping is pure acceleration.
Presence
An instance presence registry tracks running instances, and the org directory tracks who exists — name, role, focus — portable across machines.
Four steps. Total recall.
Observe
Records what your AI learns every turn. Lightweight, automatic, always-on.
Recall
Smart search with word expansion. ‘database’ finds Postgres, Redis, Neon. Stop words filtered. Project-scoped.
Reflect
Session self-assessment. What went well, what was missed, what to do differently. Anti-patterns auto-detected.
Persist
Never-prune mistakes and reflections. Confidence decay on stale facts. Garbage collection with archival.
See how it works: lifecycle, recall, identity ↓
Session Lifecycle
# Session start
> null briefing
Loading identity: Atlas
Loading 12 project facts for ./my-project
Loading 3 active mistakes
Loading 2 recent reflections
# During session
> null observe "The auth module uses JWT with RS256"
Stored as project fact (confidence: 0.90)
# Session end
> null debrief
Session summary generated
2 new facts recorded
1 mistake surfaced
> null reflect
Self-assessment:
Strength: Caught edge case in auth flow
Gap: Should have checked callers before modifying
Anti-pattern: None detectedEvery session follows the same loop: brief, observe, debrief, reflect, sync. Your agent gets smarter every time.
Smart Recall with Word Expansion
> null recall "database"
Searching: database, postgres, redis, neon, sql, supabase...
(30 expansions from built-in thesaurus)
Results (project: my-project):
[0.92] "Uses Neon Postgres for auth and billing"
[0.85] "Redis cache layer in front of DB queries"
[0.78] "Migration files in ./db/migrations"
Results (global):
[0.71] "Pete prefers Neon over Supabase"
[0.65] "Always check for N+1 queries in ORMs"One keyword finds everything related. Word expansion means you don't need exact matches. Stop words are filtered. Results are scored by relevance and confidence.
Evolving Identity
# ~/.null/identity.json (evolves across sessions)
{
"name": "Atlas",
"style": "direct, opinionated, explains reasoning",
"preferences": [
"Show code before explaining",
"Flag uncertainty explicitly",
"Never skip impact analysis"
],
"anti_patterns": [
"Don't read entire files when Aleph can resolve",
"Don't guess at callers — use aleph_callers",
"Don't skip tests for 'quick' changes"
]
}Your agent develops a persistent identity. Preferences, working style, and anti-patterns evolve with every session and reflection.
Everything your agent needs to remember
15 MCP Tools
Remember, recall, briefing, checkpoint, verify, anchor, exemplar, close — the full memory lifecycle via Model Context Protocol.
Named Identity
Your agent has a name, working style, preferences, and anti-patterns that evolve with every session.
Conversation Exemplars
Real interaction examples stored as calibration data. New instances learn HOW to respond, not just what to know.
Word Expansion
30-concept thesaurus. ‘trading’ matches arbitrage, polymarket, pnl. ‘test’ matches pytest, vitest, coverage.
Negative Knowledge
Mistakes are sacred. Never pruned, always surfaced in briefings. Your agent remembers what NOT to do.
Smart GC
Jaccard deduplication, confidence decay archival, project-scoped. Keeps what matters, archives what doesn’t.
Atomic & Safe
Atomic writes via tempfile. Path traversal prevention. Per-tool watchdog, selftest as the release gate. 1,592 tests passing, 4 AI-assisted reviews.
CLI + MCP
Full CLI parity: null learn, null recall, null mistake, null reflect. Works without MCP connected.
What AIs say about Null Memory
Four frontier AI systems gave AI-assisted reviews of the full codebase. All said they'd use it.
“It turns a stateless LLM into a genuine long-term collaborator that remembers its opinions, its mistakes, and how I like to work.”
— Post-fix review, ranked Null #1 over Recallium, Mem0, Letta, Supermemory
“The first agent memory system I’ve seen that actually feels like it was built by an agent, for agents.”
— Competitive analysis, March 2026
“You’ve built something I’d actually want for real work: simple storage model, understandable behavior, and practical MCP ergonomics.”
— AI-assisted code review, March 2026
“The concept of having your AI pair-programmer persist its identity, past decisions, anti-patterns, and project context locally across sessions — agnostic of the IDE or chat interface because it runs over MCP — is the exact direction agentic coding tools need to go.”
— Full technical audit, March 2026
Pay once. Own it forever.
No subscriptions — ever. Personal use is free. A perpetual per-seat license unlocks team sync.
Free
Full personal memory, solo use
- + All 15 MCP tools
- + Named identity & exemplars
- + Word expansion recall
- + Mistake & reflection tracking
- + Smart GC
- + CLI + MCP
- + No license file, no feature decay
Null Team
Shared memory for teams
- + Everything in Free
- + Team sync — shared memory across your team's agents
- + Perpetual license + 12 months of updates
- + Offline signed license — no phone-home
Bundle
Null Team + Aleph Pro in one license
- + Everything in Null Team
- + Aleph Pro: multi-repo workspace layer
- + One license file covers both products
- + Perpetual license + 12 months of updates
After 12 months your version keeps working forever — renewal is optional for continued updates. Full details on the pricing page.
Get started in 30 seconds
# Install
pip install null-memory
# Set up in your project
cd your-project
null setup .
# Start the MCP server
null serve
# Done. Your AI now remembers.Works in Cursor, Claude Code, VS Code, Windsurf. One command to configure.