Developer Preview Ferrosa Memory — structured, linked, auditable memory for agentic systems.

Ferrosa Memory

MCP-native memory server

Knowledge that survives the session, and the model.

A session ends and its context goes with it. The next one rebuilds the same understanding from scratch, at the same token cost, and arrives at a slightly different answer.

Ferrosa Memory is where that understanding lives instead — typed, bi-temporal, exportable, and open source, running next to your agent, as a shared service, or in a container.

SPECIMEN agent_memory.temporal_events
fact_text
“deferring the retry backoff to a later pass”
entity
ferrosa-dbaas / SnapshotScheduler
event_time
2026-06-02T14:09Z
valid_until
—  STILL OPEN
Eleven weeks and four sessions later, still queryable and still attached to the entity it belongs to. A decision an agent made about your project is a fact about your project — and it is the kind that scrolls out of a chat log.

A thin memory layer over database, graph, and inference.

Agents talk to Ferrosa Memory through a small set of MCP tools — ingest, retrieve, link, explain, and forget. The memory layer owns memory semantics: entities, folds, temporal facts, retrieval outcomes, intentions, skills, Datalog rules, and explanations.

Underneath, Ferrosa Database provides CQL storage, vector indexes, graph traversal, and durable S3-backed tiering. A background "dream cycle" consolidates new memories — discovering connections, scoring importance, and decaying stale recall — while you sleep.

Read the full architecture →

Caller
Agent / IDE / Runtime
Issues MCP tool calls; carries no storage knowledge.
▼ MCP tools
Interface
MCP tool surface
ingest / retrieve / link / explain / forget
Memory semantics
Ferrosa Memory service
entities + typed graph bi-temporal facts context segments + folds hybrid retrieval (RRF) dream-cycle consolidation Datalog-derived facts auditable forget journal
Substrate
Ferrosa Database
CQL storage HNSW vectors property graph SPARQL SQL engine Bolt protocol S3 tiering
Tool catalog

Two tiers: a focused default set, and the full toolbox.

Agents see a focused Tier 1 set of everyday tools by default. The full Tier 2 surface — batch operations, folds, bi-temporal facts, memo & plan tracking, skills, consolidation, and the Datalog governance plane — unlocks with include_all when an operator needs it.

Tier 1 — Everyday memory (21 tools)

Ingest and recall: ingest, search, find, edge, chunk & turn context, intention checks, the full session-task surface, feedback, stats, configure, and forget. Everything an agent reaches for in a normal session.

Tier 2 — Advanced & operator (58 tools)

Batch ingest/edit, trajectory folds, bi-temporal facts, memo & plans, the intention lifecycle, stored skills, dream-cycle consolidation & graph inference, the Datalog expert system (rules, claims, approvals, explanations), and config/introspection.

Browse the full tool catalog by level →

Recall you can trust, and memory you can remove.

forget

Auditable forgetting

Propose candidates, see the blast radius (edges, temporal chains, derived facts), then confirm a reversible retraction or a hard delete — all recorded in an append-only forget journal with crash recovery.

recall

Recall-relevance hardening

Source-aware relevance guards, lexical-overlap checks, authority/PageRank adjustments, and workspace filtering so hooks prefer silence over noisy context. Native full-text indexing speeds lexical recall.

session

Durable session tasks

First-class task state — focus stack, working set, and recovery hints — survives restarts and supports multi-agent handoff, so an interrupted agent can resume where it left off.

hooks

Workspace-isolated hooks

Session-start recall derives stable, workspace-specific sessions instead of leaking process-global state, and fallback recall stays semantic/procedural — never raw episodic transcript.

Run it locally or self-hosted; keep the workbench and visualization surfaces operator-only unless you configure TLS, auth, and tenant isolation.
How it compares

Built for structure, not just similarity.

Vector stores answer “what looks similar?” That is a useful signal and a poor ranking. In Ferrosa Memory similarity is three of seventeen weighted channels — lexical BM25 and full-text carry exactly as much weight, alongside phonetic matching, graph authority, workspace affinity and recency.

Capability Ferrosa Memory Vector store + RAG
Typed knowledge graphYesNo — similarity only
Bi-temporal fact historyYesOverwrite / append
Hybrid retrieval (lexical + vector + phonetic + graph)Yes — RRF over 17 weighted channelsVector (± keyword)
Explainable derived factsYes — DatalogNo
Auditable forgettingYes — blast radius + journalManual delete
Automatic consolidationYes — dream cycleNo
MCP-native toolsYesFramework-dependent
Research foundation

Engineering informed by the memory literature.

Ferrosa Memory is product engineering, not a paper implementation — but the design draws on recent and classic work on agent memory, virtual context, graph retrieval, prompt compression, Datalog-style inference, and memory evaluation.

Selected references from the local research corpus. Links point to canonical arXiv records.
PaperarXivTheme
LoCoMo: Evaluating Very Long-Term Conversational Memory of LLM Agents2402.17753long-term conversational memory
LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory2410.10813memory evaluation
Memory in Large Language Models: Mechanisms, Evaluation and Evolution2509.18868memory taxonomy
Episodic Memory is the Missing Piece for Long-Term LLM Agents2502.06975episodic memory
Microstructures and Accuracy of Graph Recall by Large Language Models2402.11821graph recall
From Local to Global: A Graph RAG Approach to Query-Focused Summarization2404.16130GraphRAG
LongLLMLingua: Accelerating and Enhancing LLMs in Long Context Scenarios via Prompt Compression2310.06839long-context compression
InfLLM: Training-Free Long-Context Extrapolation for LLMs with an Efficient Context Memory2402.04617context memory

Give your agents memory worth keeping.

Start local, inspect memory in the workbench, and watch raw context become linked, durable, queryable memory. Self-hosted today — from a single workgroup machine to a replicated cluster on the distributed Ferrosa store. A managed service is on the roadmap.

How it works Set up locally