Loading ToolBench
A hippocampus for Claude Code: an MCP server that scores each memory's surprise at write time and ranks recall by surprise·decay + cosine, so old-but-pivotal items stay visible. Pure Rust, SHODH-compatible SQLite schema (drop-in DB swap with mcp-memory-service-rs), Apache-2.0 OR MIT.
claude-hippo has well-structured schemas and descriptions for most tools, but suffers from naming inconsistency (duplicate aliases), overly technical descriptions that don't guide LLM selection, and missing output schema documentation. The memory tool is complex with many optional parameters that lack clear mutual-exclusivity guidance. Error handling is present but not recovery-focused. Descriptions average ~150 chars but often bury key details in technical jargon rather than stating WHAT/WHEN/WHY.
Alias for hippo_forget. SHODH-compatible tool name for deleting memories.
Soft-delete a memory by content_hash or id. Dry-run: report match without deleting.
List the N most recent memories.
Semantic search + surprise-weighted ranking. Top-N results to return. Disable surprise-weighted ranking (pure cosine similarity only). Per-call override for the KNN over-fetch multiplier. Larger values give surprise rerank a wider candidate pool at the cost of more SQL work. Default = server-wide setting (6 in v0.3, was 3 in v0.2). Set to `limit / 1` to disable over-fetch entirely. Caps at 1 minimum. v0.5 Phase B: recall mode. `"semantic"` (default) is the v0.4 behavior. `"associative"` follows Hebbian edges from a seed (semantic top-1 if `seed_id` is absent). `"hybrid"` merges semantic hits with associative neighbors, deduped, ranked by combined score. v0.5 Phase B: explicit seed for `mode == "associative"` or `"hybrid"`. When unset, the seed is the top-1 semantic match for `query`. Useful for clients that already have an id from a prior recall and want to expand the local subgraph without re-running semantic search.
The memory content to store. Tags as array of strings (e.g. ["auth", "decision"]). SHODH MemoryType (Decision / Learning / Discovery / Pattern / etc). Defaults to "Observation". User-marked importance, 0.0..=1.0. Increases the surprise score. Free-form metadata stored as JSON. Reserved namespace `_hippo`.
Duplicate tool aliases (store_memory, retrieve_memory, list_memories, delete_memory) create cognitive load and force LLMs to reason about which variant to use. Aliases should be removed or clearly marked as deprecated.
Descriptions are overly technical and bury actionable guidance. E.g., hippo_recall description leads with 'Semantic search + surprise-weighted ranking' rather than 'Find memories most relevant to your query, ranked by importance and recency.' LLMs need WHAT/WHEN/WHY, not implementation details.
Output schemas are not documented. Tools return results but LLMs cannot see the structure of returned fields (e.g., what fields does hippo_recall return? Is it {id, content, tags, surprise_score}?). This forces LLMs to guess and risks extraction errors.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 50 | <=2025-11-25 | v2 |
Compact summary of the most recent session's memories, ranked by surprise.
Alias for hippo_list_recent. SHODH-compatible tool name for listing recent memories.
Anthropic Memory Tool compatibility layer. Unified MCP tool supporting commands: view, create, str_replace, insert, delete, rename. Operand path (most commands). Must start with `/memories`. `view`-only: optional `[start, end]` 1-indexed line range. Out of range silently clamps. `create`-only: full file text to store. `str_replace`-only: substring to find (must be unique in the file). `str_replace`-only: replacement string. `insert`-only: 0-indexed line after which to insert (0 = before line 1). `insert`-only: text to insert. `rename`-only: source path. `rename`-only: destination path. claude-hippo extension: tags to attach to the stored memory. Not part of Anthropic's spec but harmless if unused.
Health probe returning vec_version, memory_count, and uptime.
Alias for hippo_recall. SHODH-compatible tool name for semantic search with surprise-weighted ranking.
Alias for hippo_remember. SHODH-compatible tool name for storing memory content with surprise scoring.
memory tool has 10 optional parameters with complex mutual exclusivity (e.g., old_str/new_str only valid for str_replace, insert_line/insert_text only for insert). Descriptions do not clearly state which params apply to which commands, risking invalid invocations.
ping tool is deprecated in MCP spec (removed as of 2026-07-28). Should be removed or replaced with a health check that returns structured status (e.g., {status: 'healthy', uptime_seconds: 12345, memory_count: 500}).