Semantic retrieval and memory for your LLM.

The amnesia problem

"Good morning! How can I help you today?"

DAY 1DAY 1DAY 1 DAY 1DAY 1

Every AI session wakes up blank.
Yesterday never happened.

The cost of forgetting

0

Re-explain. Re-paste. Re-burn tokens — every session.

The broken fix

The same AI that reads your memory
rewrites it. Your rules decay.

The shift

INTELLIGENCEthe LLM uses the memory

STORAGEREKALL protects the memory

REKALL separates storage from intelligence.

Memory infrastructure for LLM agents

Capture

Sessions saved as deduplicated, verifiable chunks — with contradiction checks and provenance.

Survive

Working memory that lives through context-window compaction. Resume mid-task, not amnesic.

Rekall

Hybrid search that finds the right chunk — semantic, keyword, temporal, scoped.

The daily loop

/startrestore your context + related sessions

➞

worknotes captured as you go

➞

worksurvives the context wipe

➞

/savesynthesised into memory

☾ Background schedulers nightly · UTC

Dream consolidation 02:00 Confidence decay 03:00 Missing knowledge 04:00 Synonym gaps 05:00

The proof

0%

Technical-query success

vs 0% for vector-only search

~500

Tokens per Rekall

not 27,000

60-day

Confidence half-life

memory that decays like real knowledge

Why it's different

✕Not a wiki — rewrites dilute your rules.

✕Not a knowledge graph — no reasoning, no working memory.

✓Memory that behaves like memory.

Built to last

REKALLSELF-HOSTED

Claude Code

OpenCode

ChatGPT

Gemini CLI

PostgreSQL + pgvector  ·  HTTP / MCP  ·  your data never leaves

github.com/cassiodias/rekall

Coming soon