Durable recall
Indexes files, docs, handovers, session logs, project facts, and selected corpora into a refreshable memory layer.
Persistent memory + factual control
Remanentia
Evidence memory for local AI systems that need to remember operational context, retrieve sources, and verify answers before they leave the system.
The memory layer of the Anulum agent runtime — Remember · Verify · Connect · Contain.
100%current operational recall
75.8%factuality benchmark BA
72.2%LongMemEval R11 (oracle)
3,093current repo test count
What it does
Turns scattered project knowledge into controlled evidence.
Remanentia is built for operators who need AI systems to use memory without turning every internal record into public text.
Evidence search
Combines keyword search, embeddings, reranking, compiled facts, and direct recall paths for source-grounded answers.
Factual guard
Director Class AI verifies generated answers against retrieved evidence and blocks unsupported public output.
Public safety
Public APIs expose allowlisted corpora and redacted snippets while private operational memory remains internal by default.
Fleet feed ingest
Ingests findings and decisions from the SYNAPSE coordination feed through an admission gate, so the fleet's recorded work becomes durable, queryable memory.
Operating loop
Retrieve first. Verify next. Answer last.
Remanentia separates memory from output. Retrieval produces evidence; the control layer checks that evidence against the intended answer and deployment policy.
1Request enters from an agent, product, or API path.
2Policy selects private recall, public search, or no-generation mode.
3Memory returns sources, snippets, and compiled facts.
4Factual checks and public-output controls accept or reject the answer.
Measured state
Proof points from the current local operator build.
Recall@1 / @5 / @10 / @20 on the current operational suite.
60,890 chunks, 768 dimensions, full rebuild measured at 391.06s.
p50 through the local API path, with p95 measured at 121.4ms.
Balanced accuracy on the factual-consistency benchmark.
LongMemEval R11 temporal score, with 72.2% overall (oracle setting).
Current Remanentia repository test count, verified by a collection-only run.
Part of the Anulum agent runtime
Four layers a model can't supply on its own.
Remanentia is the memory layer of a four-part runtime. Each product stands alone; together they run an agent that remembers, checks itself, coordinates with its peers, and contains what it executes.
Remanentia
Durable, source-grounded memory with evidence search and factual control — this site.
Director-AI
Checks what the agent says and does; high-impact actions escalate to human approval.
Explore Director-AI →Synapse Channel
Local-first coordination bus with a durable event log agents can replay, query for causality, and prove against a Merkle root.
Explore Synapse →HushLine
A local guard for the commands an agent runs — secret redaction, output caps, directory gates.
Explore HushLine →Deployment fit
For teams building memory-backed AI products on private knowledge.
Self-hosted memory service
Run Remanentia beside existing agents and applications, with local-first inference and explicit hosted fallback modes.
Controlled public API
Expose only approved corpora, verified snippets, and factual-control decisions to customer-facing surfaces.
Commercial factual guard
Use Director Class AI licensing when the verification layer ships inside closed-source products or SaaS systems.