Remanentia
Persistent memory for AI systems. Indexes your docs, logs, handovers, and project facts into durable, source-grounded recall with vector search and entity graphs.
remanentia.com →Remember · Verify · Connect · Contain
Raw models forget, make things up, work alone — and leak when they run.
Four layers a language model can't supply on its own: durable memory, output it can verify, a way to coordinate with other agents, and a safe boundary for the commands it runs. Remanentia, Director-AI, Synapse Channel, and HushLine add exactly those.
4runtime layers, one stack
72.2%LongMemEval memory recall
FreeDirector-Lite entry tier
MCPnative across the suite
The three layers
One runtime. Four jobs a model can't do alone.
Each layer is its own product you can adopt on its own — or run together for an agent that remembers, checks itself, talks to its peers, and runs commands without leaking.
Director-AI
A guardrail that checks what the agent says and does. Response-level fact-checking against retrieved evidence, plus action control that reviews high-impact tool calls before they run.
Explore Director-AI →Synapse Channel
A live channel where autonomous agents message, hand off work, and coordinate across projects and terminals — one bootstrap command to join the bus.
Request early access →HushLine
A local guard for the commands an agent runs. Redacts secrets from tool output, caps noisy output before it breaks the parser, and gates which directories may execute — no network, no telemetry.
Request early access →Director-AI, three editions
Start free. Grow into output and action control.
Director-AI scales from a three-line guard you drop into any project to a commercial layer that gates an autonomous agent's real-world actions.
Director-Lite
Three-line guard facade with a model-free heuristic default and facts/RAG handoff. Optional NLI upgrade. pip install director-ai-lite
Director-AI
Full response-level hallucination guardrail: NLI + RAG fact-checking with token-level streaming halt before an unsupported answer ships.
Director-Class-AI
Runtime action control. Reviews high-impact shell, SQL, infrastructure, API, and MCP actions before dispatch; routes uncertain or high-risk ones to human approval.
Working together
Route. Recall. Draft. Verify.
The layers compose into one request path. Memory grounds the answer, the guardrail checks it, and the channel moves work between agents — each layer optional, the order consistent.
1Synapse routes the request to the right agent and carries hand-offs between them.
2Remanentia returns the sources, facts, and history the agent needs to answer.
3The agent drafts an answer or a proposed action.
4Director-AI verifies it against the evidence and gates risky actions for approval before anything ships.
Adopt on your terms
Each product runs on its own. Together they run the agent.
Self-hosted, local-first
Run the memory and guardrail layers beside your own agents with local inference and explicit hosted fallback modes.
Open source, then commercial
Director-Lite and the Remanentia package are public on PyPI. Action control and the closed-source factual guard ship under commercial licensing.
MCP-native
Memory, verification, and the agent channel speak MCP, so they slot into existing agent stacks without bespoke glue.