Jeli
Sovereign, cryptographically-attested personal memory
signed at write, audited end-to-end, lives on your hardware
What it is
Jeli is a personal memory framework — a sovereign store of facts, decisions, preferences, and history that an AI assistant can read from and contribute to.
- Sovereign — lives on your hardware, under your keys. Not in a SaaS database somewhere.
- Cryptographically attested — every entry is signed at write time. Tampering and forgery are detectable.
The model: your AI assistant queries Jeli before answering, augmenting the model's context with facts only you have. New facts learned during a conversation get proposed back to Jeli for review and signing — humans-in-the-loop on memory persistence.
Design principles
The same principles that shape LegionForge shape Jeli:
- Local-first. Memory lives on your machine. Replication is opt-in and explicit.
- Cryptographic provenance. Every entry has a signature. Memory can be audited and replayed.
- Human gates on writes. New memories are proposed by the assistant and committed by you. Same pattern as HITL on destructive tool calls.
- Replace AI with determinism wherever possible. Memory retrieval is structured (tags, types, dates) before falling back to embedding search.
Use cases
- Personal AI assistants that need stable, user-controlled context
- Long-running agents that should accumulate knowledge across sessions
- Compliance scenarios where memory provenance has to be auditable
Integration with LegionForge
Jeli exposes an MCP server. The LegionForge framework registers Jeli as a context source and gains read / propose access through the standard MCP path. The HITL approval gate in LegionForge handles the "approve this proposed memory" step.
Status
Active. Public. See the GitHub repo for the latest.