Built by Prashant Malik, co-creator of Apache Cassandra.Scattered experienceShared memoryDecision and action
CortexDB memory
Waiting?
ScatteredReconciledDecisionLearns
Outcome returns as new experience ↩
Memory infrastructure for AI agents
Scoped, time-aware context for your agents. Audit trails you can query.
- 93.8%LongMemEval-S · 469 / 500
- 91.8%LoCoMo
- 48msp50 recall
docker run -d -p 3141:3141 -v cortexdb-data:/data -e OPENAI_API_KEY=$OPENAI_API_KEY -e ANTHROPIC_API_KEY=$ANTHROPIC_API_KEY cortexdb/cortexdb:latest
Teams building on CortexDB
Leena AI
LimeChat
RISA Labs
TinyHumans
StatsUSA
theStacc
Leena AI
LimeChat
RISA Labs
TinyHumans
StatsUSA
theStacc
Built by Prashant Malik, co-creator of Apache Cassandra.
Watch scattered experience become usable memory.
Three implementation stories show the complete loop: source experience, evolving memory, the decision it enables, and the outcome that returns.
03 / Memory in motionStory plays in view
The experience already exists
Your systems remember in fragments.
Tickets, conversations, documents, product events, decisions, corrections, and outcomes live in tools that were never designed to learn together.
Issue resurfaced after the latest change.Previous action stored elsewhere
The requirement changed.Correction isolated in chat
Escalate when both conditions are true.Rationale detached from outcome
Relationship state changed.Updated yesterday
Current approval threshold.Supersedes v4.1
Workflow completed after revision.Outcome isolated
Evidence and timing considered.Why is stored elsewhere
Escalation resolved the issue.Next agent has not learned this
Valid
14 AugRequirement changed · evidence attached
14 AugRequirement changed · evidence attached
Learned
15 AugPolicy v4.2 supersedes v4.1
15 AugPolicy v4.2 supersedes v4.1
Outcome
16 AugEscalation resolved the issue
16 AugEscalation resolved the issue
What should happen next?
CurrentRequirement changed after the original case.
Valid nowThe revised policy governs this decision.
Prior outcomeEscalation resolved the same condition.
Outcome returns as experience
Memory state updated
Memory state updated
Memory has more than one job.
Persistence, changing truth, experience, shared context, and selection are different behaviours—not five labels for the same retrieval trick.
Active memory behaviour
Carry the relationship across sessions.
Preferences, interactions, and feedback stay available when the user returns and when another agent continues the work.
If memory has five jobs, why is similarity alone insufficient?
Session 01 · preference observed
Memory · preference + source retained
Session 08 · context available
Policy v4.1 · valid until 14 Aug
Correction learned on 15 Aug
Policy v4.2 · valid from 14 Aug
Action · escalation chosen
Trace · reason + evidence
Outcome · issue resolved
Agent A · learns a constraint
Shared memory · source intact
Agent B · uses it later
Many plausible memories
Current task + state
Only decision-relevant context
The closest memory can still be wrong.
CortexDB does not stop at semantic similarity. It checks current state, validity, revisions, relationships, and evidence before context reaches the agent.
Decision nowWhich approval rule governs this request?
Nearest semantic matchManager approval above the old threshold.
SupersededCurrent decision contextRevised threshold applies from 14 August.
Valid nowRelated precedentEscalation resolved the same condition.
OutcomeSupporting sourceCorrection recorded on 15 August.
EvidenceDecision
gate
gate
Context selected for the agent
The valid memory wins.
The highest similarity score is rejected because the policy was superseded. The current rule arrives with its source and revision history intact.
Selected memoryPolicy v4.2 · valid from 14 Aug
Recorded 15 Aug · supersedes v4.1 · evidence attached
Everything you need to compound intelligence.
One memory substrate for preserving experience, tracking changing truth, selecting context, surviving failure, sharing safely, and connecting the systems where work happens.
One model · five addressable layers
Preserve experience before you compress it.
Capture immutable events, form episodes, derive facts and beliefs, and synthesize understanding. Read each layer independently or retrieve them as one stratified context pack.
Events → Episodes → Facts → Beliefs → Understanding
Events · lossless captureEpisodes · sealed spansFacts · temporal recordsBeliefs · confidence attachedUnderstanding · synthesized concepts
Valid time + recorded time
Know what was true and when you learned it.
Every derived record carries two clocks. Corrections update the present without deleting the state an earlier decision actually saw.
Ask: what did the system believe on 15 April, as known on 1 May?
Valid time →Recorded time →v4.1 recordedChange validv4.2 learned
Hybrid retrieval · one round trip
Recall the right mix, not one flat bag.
Fuse semantic and lexical retrieval, rerank candidates, and return events, episodes, facts, beliefs, and understanding with narrative context and provenance.
HNSW + BM25 + RRF fusion + optional reranking
EventsEpisodesFactsBeliefsUnderstanding
Durability before derivation
Keep memory through crashes and replay.
Append captures to an immutable write-ahead log, persist them in RocksDB, and reproduce derived layers asynchronously instead of rewriting memory on the write path.
WAL → RocksDB → Raft replication
Immutable
WAL→RocksDB
storage→Replicated
cluster
WAL→RocksDB
storage→Replicated
cluster
One hierarchy · policies at every scope
Share memory without flattening boundaries.
Organise tenants, teams, agents, and users in hierarchical scope paths. Recall up for shared context or down for specific context while each scope retains its own policy.
org:acme / dept:eng / user:alice
org:acmedept:engineeringuser:alice · agent:support
Experience already lives in tools
Bring source context into one memory layer.
Connect workplace systems, conversations, documents, and product data while retaining the source needed to audit important memories later.
16 built-in connectors reported in the current product material
SlackGitHubJiraSalesforceGmailNotionDocumentsProduct events
The memory layer, not another memory bag
Keep the controls production AI needs.
Make preservation, deletion, authorization, and audit part of the memory model instead of rebuilding them around a vector store in application code.
Designed as infrastructure: inspectable, replayable, scoped, and auditable
Data preservationImmutable events; derived layers reproducible by replay
Write pathNo LLM required on every write
ForgettingSelective forget with preview before execution
AuthorizationPASETO tokens and tiered capabilities
AuditChained records with tamper verification
Fast recall on a durable memory substrate.
The recall path plans, searches, reranks, gates, and enriches context. Under it, a Raft-replicated RocksDB cluster keeps the memory service available and horizontally scalable.
Cognitive recall pipeline
From query to decision-ready context.
Hybrid search, reranking, validity gates, and graph enrichment in one recall path.
01Query
02Planner
034-CH search
04Reranking
05Gate
06KG enrich
Distributed cluster · Raft consensus
Memory that survives the process.
Immutable capture, replicated storage, leader election, and horizontal scaling beneath every recall.
Healthy · RF 3LeaderN2N3N4N5
Measured memory, not a demo that remembers.
Published here as reported runs, with the test setup visible. Benchmark results, latency, and cost belong together because accuracy without operating context is incomplete.
LongMemEval-S · server-parity run
Long-term memory accuracy
93.8%469 / 500
Supermemory
98.6
AgentMemory
96.2
Mastra
94.9
CortexDB
93.8
Mem0
93.4
Hindsight
91.4
Zep / Graphiti
71.2
Reported production-server run: WAL + RocksDB + HNSW. Claude Opus 4.6 answerer. Run dated 2 May 2026.
LoCoMo · categories 1–4
Long-conversation memory
91.8%Latest result
Single-hop91.6%
Temporal87.9%
Multi-hop79.8%
Open-domain64.6%
Reported stage: cats1234_typed_tokens_v1, judged 3 May 2026. Mem0’s cited self-reported score is 91.6%; CortexDB trails on this benchmark while leading it on the reported LongMemEval-S run.
End-to-end answer profileReported on the 93.8% LongMemEval-S run
48msp50 recall
2.5sp50 generation
9.7sp50 total
$0.12per question
Memory should fit your infrastructure, not escape it.
Define where memory runs, what crosses data boundaries, and who operates each layer before recommending an architecture.
Boundary 01Where must memory run?
Boundary 02What can leave the data boundary?
Boundary 03Who operates each part?
Plan deployment around your infrastructure, data-boundary, and operating requirements.
Discuss deployment requirements ↗
10 / Start
Start where history changes the answer.
Bring one agent workflow. We will map what it needs to remember, what changes over time, and where the memory layer must run.
Discuss an agent workflow ↗Bring one workflow where history or sequence changes the decision.