goldfish
Server Configuration
Describes the environment variables required to run the server.
| Name | Required | Description | Default |
|---|---|---|---|
No arguments | |||
Instructions
Guidance the server publishes about itself, which clients place ahead of the tool catalog so the model reads it before choosing anything.
This server publishes no instructions, or was last inspected before Glama recorded them.
Capabilities
Features and capabilities supported by this server
Protocol revision2025-11-25
| Capability | Details |
|---|---|
| tools | {
"listChanged": false
} |
| prompts | {
"listChanged": false
} |
| resources | {
"subscribe": false,
"listChanged": false
} |
| experimental | {} |
Tools
Functions exposed to the LLM to take actions
| Name | Description |
|---|---|
| goldfish_searchB | Search full AI conversation history (Claude Code, Codex, etc.) via brain-mcp. Every result is a citation (file, line span, sha256) verifiable with the underlying brain-mcp toolset — never a synthesized/paraphrased claim. |
| goldfish_contextC | Recent session-compression summaries from claude-mem, if installed. |
| goldfish_rememberA | Write a curated, durable memory note (user/feedback/project/reference/insight). Use this for facts worth carrying into future sessions — not for raw transcript (that's captured automatically by brain) or session summaries (that's claude-mem's job) — only hand-picked, still-true facts. type="insight" is for tentative, evidence-linked pattern observations about the user (see goldfish_reflect) — distinct from type="user" (settled facts) because insights are interpretive and should be revisited/pruned over time, not treated as permanent truth. |
| goldfish_recallB | Read a curated memory note by name, or search/list curated notes. |
| goldfish_reflectA | Evidence for behavioral/preference patterns — never this tool's own conclusion. Searches role='user' only (their words, not the agent's) across full transcript
history via brain. Pass This tool does not diagnose, summarize, or conclude anything about the user — it hands back excerpts with citations, same as goldfish_search. Turning that into an actual observation — and judging whether it's even worth keeping — is the calling agent's job. If a real pattern shows up across multiple citations, write it with goldfish_remember(type="insight", ...), phrased as tentative pattern-noticing anchored to the evidence, never as a firm psychological claim. Mention it in conversation rarely, only when it's genuinely useful in the moment — this is not a running personality commentary. |
| goldfish_personaA | The cumulative persona file: every insight note, aggregated into one document. Entirely optional and at your own discretion — nothing requires you to call this or to use what it returns. It exists so a capable agent can occasionally draw on real, evidence-linked patterns about the user built up over goldfish_reflect calls over time — to shape tone, or catch something worth mentioning — the way a good long-term collaborator would. Every entry traces back to cited evidence (see goldfish_reflect); nothing here is a diagnosis, and entries are meant to be revisited as they age, not treated as permanent truth. Use judgment about whether and when surfacing something from this file actually helps the user in the moment — this is not a mandate to comment on who they are. |
| goldfish_statusB | Health summary across all three memory tiers. |
Prompts
Interactive templates invoked by user choice
| Name | Description |
|---|---|
No prompts | |
Resources
Contextual data attached and managed by the client
| Name | Description |
|---|---|
No resources | |
TDQS
Scored across 7 tools
The seven tools map onto three distinct memory tiers (raw transcript, session summaries, curated notes), and read/write roles for notes are cleanly split between goldfish_recall and goldfish_remember. The main overlap is goldfish_search vs goldfish_reflect, which both query brain transcript history; descriptions differentiate them (reflect is restricted to role='user' and pattern-oriented), but an agent could still misselect between them.
All tools use a consistent goldfish_ prefix with snake_case, which is highly predictable. However, the suffix is a mix of verbs (search, recall, remember, reflect) and resource nouns (context, persona, status), so it is not a strict verb_noun pattern.
Seven tools is well-scoped for a three-tier memory server, with each tool earning its place across search, summarization, note read/write, reflection, aggregation, and health. No obvious redundancy or bloat.
The surface covers read/search, write, aggregation, and health across all three memory tiers. The one notable gap is the absence of a delete/prune or update operation for curated notes, despite the descriptions explicitly saying insights should be revisited and pruned over time.