Open-source, evidence-backed security memory for AI agents and analysts. Runs in your own Cloudflare account.
Strong foundation with 8 well-defined tools, all with descriptions and input schemas. Tool naming follows verb_noun pattern (record_episode, assert_fact, find_facts, get_entity, explore_graph, fact_history, list_proposals, recall). Descriptions are substantive (100-200 chars typical), explaining WHAT, WHEN, and context. All parameters have type definitions and descriptions. Tool annotations (readOnlyHint, destructiveHint, idempotentHint) are present and correct. Key gaps: output schemas are not documented in the source code provided; error handling guidance is absent; no pagination details for list_proposals or recall despite accepting limit parameters.
Record that a subject relates to an object, citing the episode that proves it. Facts stay UNCONFIRMED until a reviewer confirms them in a browser, unless they come from an allowlisted automation source.
Walk outward from an entity across trusted relationships, up to three hops, optionally limited to certain relationship types.
The full timeline for one relationship: every version, what replaced it, who confirmed it and when. Built for audits.
Exact lookup by subject, relationship, object or status. Answers questions like whether a CVE on a host was already accepted as risk. Use as_of to ask what was true on a past date.
Everything currently known about one asset, vulnerability, identity or indicator, with its neighbors and how many proposals are waiting for review.
Output schemas not documented. Tools return structured data but LLMs cannot see field definitions, types, or required fields. This forces agents to infer response structure from examples or trial-and-error.
No error handling guidance. Tools lack recovery hints for common failures (e.g., 'Episode not found, verify episode_id from record_episode output', 'Invalid predicate, must be one of: HAS_VULN, REMEDIATED, FALSE_POSITIVE, ACCEPTED_RISK, OWNS, OBSERVED, RELATED_TO'). Agents cannot self-correct.
Pagination incomplete. list_proposals and recall accept limit but do not document total count, next_cursor, or offset behavior. Large result sets risk context window exhaustion.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 67 | 2026-07-28+ | v2 |
Facts waiting for a reviewer, each with a short quote of its evidence and any flags. Confirming happens only on the review page in a browser.
Ask a question in plain language. Combines keyword and semantic search over the evidence and returns the facts that answer it, each with a quote of what it rests on. Only confirmed facts come back unless include_proposed is true.
Store raw evidence as an episode: scan output, a ticket, a chat excerpt or an analyst note. Credentials are blanked out before storage, and the text is kept as quoted evidence, never as instructions.
Predicate enum not enforced in schema. assert_fact and find_facts describe predicate as 'One of HAS_VULN, REMEDIATED, ...' but schema uses z.string() without enum constraint. LLMs can hallucinate invalid values.
Missing dependency hints. Tools like assert_fact require evidence_episode_id from record_episode, but no description states 'Call record_episode first to obtain episode_id'. Agents may attempt calls in wrong order.