Skip to main content
Glama
fireweed-mcp

fireweed-mcp

fireweed-mcp

Agent memory where every fact carries a receipt.

remember(claim    = "Priya joined Acme in 2019 under duress.",
         evidence = "Priya Raman joined Acme in 2019 as a logistics analyst.")

REFUSED (asserts_more_than_evidence) — the claim adds something the evidence does not say.
  claim   : Priya joined Acme in 2019 under duress.
  evidence: Priya Raman joined Acme in 2019 as a logistics analyst.
recall("Priya's salary")

ABSTAINED (unknown_predicate) — no claims ground "salary"; 1 claim about Priya Raman exists
This is a refusal, not an empty result.
forget("Priya")

ERASED Priya Raman — certificate issued
  signature            : hmac-sha256:f4d0768ef3b0fec624afec12f25bfd91…
  nodes in closure     : 1
  every probe abstains : True
  bystanders surviving : 1

That last one is the artifact behind "delete me from your agent's memory — and prove it."

Install

uvx fireweed-mcp          # try it
pip install fireweed-mcp  # keep it
claude mcp add fireweed -- uvx fireweed-mcp

No dependencies. No API keys. No model — nothing in this server calls an LLM.

Related MCP server: Nahuali

What it does

tool

remember

admits a claim only if the evidence you cite supports it. Refusals are typed and say what to fix.

recall

grounded claims with the byte range they came from; abstains and names the term it could not ground

verify_receipts

re-hash every source, re-slice every range — tamper-evident

trace_evidence

audit one memory backwards to its evidence's arrival: the bytes it binds, whether they still match, the ledger event that recorded the document, and whether the chain verifies

review_reads

what has been asked of this substrate and what it answered — off by default, and queries are salted fingerprints unless you also opt into recording text

forget

erasure with exact closure and a signed certificate; bystanders survive

export_memory

the whole substrate as a portable open-format blob

Why the refusals are the point

Most memory servers store what the model says and return what's nearest. This one adjudicates.

The rule is the model proposes, deterministic code decides. Across an RPC boundary that stops being a slogan: your agent is the proposer, and it cannot talk its way past the gate, because the gate is not a prompt. Pass a claim and the text you're quoting; pure functions check that the evidence names the subject, preserves the relation, invents no numbers, and asserts nothing the span doesn't say. What survives is stored with a byte range into the source.

Then anyone can check it afterwards — including someone who trusts neither your agent nor this server. That is the whole product.

What it does NOT do

Stated up front, because this project's last headline number turned out to be measuring nothing (see the retraction, which ships with a script that proves it):

  • The server itself does not extract memories from free text. You supply the claim and the evidence, and nothing in this server calls a model. Since 0.5.0 an optional companion, fireweed_extractor, will propose claim/evidence pairs from a transcript using a model you run — and it is never trusted: every proposal goes through the same four checks a hand-written one does. Measured across four model families, admitted yield ranged from 0% to 99.3% while every unfaithful proposal was rejected with a typed reason. One 4B model produced 46 pairs its own cited span did not support; all 46 were refused. The failure mode is fewer memories, never false ones.

  • It does not make an LLM truthful. It governs what enters the record and what can be proven about it. Your model can still say whatever it likes in its own prose.

  • Recall is the weak half, and the honest number is far worse than this page used to claim. A previous version of this README said the gate finds a stored fact 98.4% of the time. That figure is withdrawn. It was measured on a corpus whose fourteen question phrasings all have a matching entry in the hand-written category table that answers them — because those entries were derived from that same corpus's failures. It measured the table's coverage of one question set, not the system's recall.

    Measured 2026-08-27 against a corpus held out on both axes — unseen personas and, crucially, unseen question phrasings:

    asked with…

    default install refuses

    the phrasings the table was built from

    4.8%

    phrasings it has never seen

    99.2%

    A default install answers almost nothing phrased in words nobody tuned for. That is the number that describes the system, and it replaces every recall claim this page previously made.

  • What is genuinely strong is the other axis. On absent-answer traps the gate correctly refuses 96.1% — it is far better at declining than at answering, and it does not fabricate. If you need a memory that never invents, this is that. If you need one that reliably finds things, it is not there yet, and the number above is why.

  • It does not yet handle multi-subject questions with scope. Questions naming exactly one subject are scoped to that subject; questions naming two or more still match against the whole store.

    Numbers come from a calibrated instrument that prints its own controls before measuring. The corpora and method live in the private evaluation repo, so treat these as reported rather than independently checkable — the write path, receipts, provenance and erasure are the parts you can verify yourself with the commands above.

Your data

~/.fireweed/mcp/ (FIREWEED_MCP_STORE to change). The substrate is an open format — see open_format/SPEC.md — and open_format/reference_reader.py reads it with the standard library alone. Your memory outlives this server, this engine, and any model. A test asserts that round trip.

Do not install fireweed-mcp[semantic]. It enables paraphrase matching in recall, and measured against the absent-answer traps it collapses correct refusal from 96.1% to 32.8% — it answers two thirds of questions whose answer is simply not in the store. A threshold sweep found no setting where it buys recall without that cost: tightened far enough to be safe, it contributes nothing at all. It stays installable because the mechanism may be salvageable when scoped to a subject's own predicates, which is untested. Until then it is off, and memory_stats tells you which mode you are in.

License

FSL-1.1-ALv2 — source-available. Free for everything except building a competing product; converts to Apache 2.0 on 2028-01-01. Full text in LICENSE.md.

Want to use Fireweed in a commercial product or competing service? → sanyamsood2@gmail.com

Available Tools

10 tools
add_sourceB

Register a source document so claims remembered against it bind to verifiable byte ranges.

Inputs: text (required), source_id (required), origin, origin_kind, +2 more.

export_memoryB

Export the whole substrate as a portable open-format blob. Readable without this server, without any model, with a stdlib-only reference reader.

Inputs: path.

forgetA

Erase everything about a subject and issue a SIGNED CERTIFICATE: exact closure, a probe battery that must all abstain, and bystanders left intact. This is the artifact for a 'delete me and prove it' request.

Inputs: subject (required).

memory_statsD

Substrate size, entities, sources held, and mode.

Inputs: no named parameters.

recallA

Search memory. Returns grounded claims with the byte ranges they came from. If the substrate cannot answer, it ABSTAINS and says which term it could not ground — treat that as a real answer, not an empty result.

Inputs: query (required).

rememberA

Commit a fact to memory. The claim is admitted ONLY if the evidence you cite supports it — you are the proposer, a deterministic gate decides. Refusals are typed and explain what to fix. Evidence must be text you are quoting verbatim, not a paraphrase.

Inputs: claim (required), evidence (required), source_id, source_text.

review_quarantineA

List claims the firewall held for review rather than storing. These were NOT written to memory; a QUARANTINE verdict means the claim could not be classified confidently, not that it was rejected.

Inputs: limit.

review_readsC

What has been asked of this substrate and what it answered. Off unless FIREWEED_MCP_READ_AUDIT=1; query text is recorded only if FIREWEED_MCP_READ_AUDIT_TEXT=1 as well, otherwise queries appear as salted fingerprints.

Inputs: limit.

trace_evidenceA

Audit BACKWARDS from a stored memory to the arrival of the evidence it rests on: the byte range it binds, whether those bytes still match, whether the document's arrival is in the append-only ledger, and whether the chain verifies. States plainly which fields are attested and which are caller-declared.

Inputs: claim (required).

verify_receiptsA

Re-hash every held source and re-slice every receipt. Tamper-evident: change one byte of a source and its receipts stop verifying.

Inputs: no named parameters.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    B3.4/5.0

    Scored across 10 tools

    Disambiguation5/5

    Each tool targets a distinct operation: writing, recalling, erasing, exporting, auditing reads, tracing provenance, verifying receipts, and reviewing quarantine. The closest pairs (review_reads vs review_quarantine, trace_evidence vs verify_receipts) are clearly separated by their descriptions.

    Naming Consistency3/5

    Most tools follow a verb_noun pattern (add_source, export_memory, verify_receipts, trace_evidence, review_reads, review_quarantine), but remember, recall, and forget are bare verbs and memory_stats is noun_noun. The style is readable and lowercase-snake throughout, but the convention is genuinely mixed.

    Tool Count5/5

    Ten tools is well-scoped for a verifiable memory substrate. Each tool covers a necessary capability—ingestion, recall, deletion with proof, export, audit, verification, quarantine review, and statistics—without redundant or filler tools.

    Completeness5/5

    The surface covers the full evidentiary memory lifecycle: add_source and remember for writing, recall for reading, forget for erasure with proof, export for portability, plus trace_evidence and verify_receipts for auditability. The append-only, evidence-bound design makes an update tool unnecessary; corrections would naturally flow through new evidence.

    Maintenance

    ActivityInactive
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      D
      maintenance
      Verified memory for AI agents — agents propose memories that are quarantined until verified against evidence, and recall() returns only trusted, fresh, and in-scope facts, preventing poisoned or hallucinated data from spreading.
      7 npm
      1
      MIT
    • A
      license
      Not graded
      quality
      B
      maintenance
      Local-first memory for AI agents with evidence-backed recall, deterministic trust verdicts, self-inspection, and a tamper-evident audit history.
      Apache 2.0
    • A
      license
      Not graded
      quality
      D
      maintenance
      An MCP server for agent memory that enables remembering, recalling, and provably forgetting knowledge with cascade deletion, cryptographic shredding, and signed erasure certificates.
      MIT