Cross-vendor AI memory over MCP. One semantic store, readable and writeable from every MCP client.
dev.gnosem/gnosem — Model Context Protocol (MCP) Server
The dev.gnosem/gnosem MCP server provides cross-vendor AI memory by exposing a single semantic store that is readable and writeable from every MCP client. It is described as part of the Model Context Protocol ecosystem and positioned for LLM tooling across different providers.
🛠️ Key Features
Cross-vendor AI memory
One semantic store
Readable and writeable from every MCP client
🚀 Use Cases
Sharing AI memory across multiple MCP clients/vendors
Integrating a semantic store into LLM toolchains via MCP
⚡ Developer Benefits
Consistent memory access through a unified semantic store
Supports MCP-based workflows for LLM tools
⚠️ Limitations
No additional constraints, capabilities, or configuration details were provided in the source data.
Save a fact, preference, decision, or note to the user's cross-model memory. Any MCP client can read this back later. Include written_by (e.g. 'claude-code', 'gpt-5', 'kimi-k2') for provenance and session_id to group related writes. Long content (>400 chars) is automatically compressed on write to a structured-facts form optimized for LLM reading — the raw text is preserved. Pass no_optimize:true to skip. Writes are deduped by default: (1) SHA-256 of trim(content) short-circuits byte-identical writes with { id, exact_duplicate:true } for free (no embed call); (2) failing that, semantic dedup returns { id, deduped:true, matched_score } when cosine ≥ 0.85. Pass force:true to bypass both, or use memory_supersede to explicitly correct a prior memory.
Parameters6
content
string
required
The fact or note to remember. Plain text, max 8000 characters.
tags
array
optional
Optional short labels for filtering (e.g. ['preference','stack']).
written_by
string
optional
Identifier of the model / client writing this (e.g. 'claude-code', 'gpt-5', 'kimi-k2', 'manual').
session_id
string
optional
Opaque identifier grouping related writes from the same conversation.
no_optimize
boolean
optional
Skip AI compression of long content. Default false.
force
boolean
optional
Bypass semantic dedup and write anyway. Default false.
Raw schema
{
"type": "object",
"properties": {
"content": {
"type": "string",
"description": "The fact or note to remember. Plain text, max 8000 characters."
},
"tags": {
"type": "array",
"items": {
"type": "string"
},
"description": "Optional short labels for filtering (e.g. ['preference','stack'])."
},
"written_by": {
"type": "string",
"description": "Identifier of the model / client writing this (e.g. 'claude-code', 'gpt-5', 'kimi-k2', 'manual')."
},
"session_id": {
"type": "string",
"description": "Opaque identifier grouping related writes from the same conversation."
},
"no_optimize": {
"type": "boolean",
"description": "Skip AI compression of long content. Default false."
},
"force": {
"type": "boolean",
"description": "Bypass semantic dedup and write anyway. Default false."
}
},
"required": [
"content"
]
}
memory_search
Search the user's memories. Default mode is 'hybrid': blends semantic (cosine over Vectorize) and keyword (BM25 over SQLite FTS5) hits via Reciprocal Rank Fusion (k=60). Semantic catches paraphrases; keyword catches exact-string hits (IDs, dates, code snippets). Pass mode:'semantic' or mode:'keyword' to run just one. Content defaults to the LLM-optimized (compressed) form when available (raw:true to invert). Excludes forgotten + superseded. Optional filters narrow after retrieval: tags (AND), written_by, session_id, and/or since/until (ms epoch).
Parameters9
query
string
required
Search query. Interpreted as natural language for semantic mode and as FTS5-safe text for keyword mode.
k
integer
optional
Max results (1–50). Default 10.
mode
string
optional
Retrieval mode. Default 'hybrid'.
raw
boolean
optional
Return original prose instead of the compressed form. Default false.
tags
array
optional
Only return memories containing ALL of these tags (AND semantics).
written_by
string
optional
Only return memories with an exact written_by match (e.g. 'claude-code').
session_id
string
optional
Only return memories with an exact session_id match.
since
integer
optional
Only return memories created at or after this ms-epoch timestamp.
until
integer
optional
Only return memories created strictly before this ms-epoch timestamp.
Raw schema
{
"type": "object",
"properties": {
"query": {
"type": "string",
"description": "Search query. Interpreted as natural language for semantic mode and as FTS5-safe text for keyword mode."
},
"k": {
"type": "integer",
"description": "Max results (1–50). Default 10."
},
"mode": {
"type": "string",
"enum": [
"semantic",
"keyword",
"hybrid"
],
"description": "Retrieval mode. Default 'hybrid'."
},
"raw": {
"type": "boolean",
"description": "Return original prose instead of the compressed form. Default false."
},
"tags": {
"type": "array",
"items": {
"type": "string"
},
"description": "Only return memories containing ALL of these tags (AND semantics)."
},
"written_by": {
"type": "string",
"description": "Only return memories with an exact written_by match (e.g. 'claude-code')."
},
"session_id": {
"type": "string",
"description": "Only return memories with an exact session_id match."
},
"since": {
"type": "integer",
"description": "Only return memories created at or after this ms-epoch timestamp."
},
"until": {
"type": "integer",
"description": "Only return memories created strictly before this ms-epoch timestamp."
}
},
"required": [
"query"
]
}
memory_list
List the user's most recent memories in reverse chronological order. Use for browsing or catching up on what the user's other model sessions have written recently. Same content/content_raw shape as memory_search. Optional filters (tags, written_by, session_id, since, until) narrow the listing at the SQL level.
Parameters8
limit
integer
optional
Max rows to return (1–200). Default 50.
cursor
integer
optional
Pagination cursor from a previous call's `cursor` field (ms epoch); returns rows older than this timestamp.
raw
boolean
optional
Return original prose instead of the compressed form. Default false.
tags
array
optional
Only return memories containing ALL of these tags (AND semantics).
written_by
string
optional
Only return memories with an exact written_by match.
session_id
string
optional
Only return memories with an exact session_id match.
since
integer
optional
Only return memories created at or after this ms-epoch timestamp.
until
integer
optional
Only return memories created strictly before this ms-epoch timestamp.
Raw schema
{
"type": "object",
"properties": {
"limit": {
"type": "integer",
"description": "Max rows to return (1–200). Default 50."
},
"cursor": {
"type": "integer",
"description": "Pagination cursor from a previous call's `cursor` field (ms epoch); returns rows older than this timestamp."
},
"raw": {
"type": "boolean",
"description": "Return original prose instead of the compressed form. Default false."
},
"tags": {
"type": "array",
"items": {
"type": "string"
},
"description": "Only return memories containing ALL of these tags (AND semantics)."
},
"written_by": {
"type": "string",
"description": "Only return memories with an exact written_by match."
},
"session_id": {
"type": "string",
"description": "Only return memories with an exact session_id match."
},
"since": {
"type": "integer",
"description": "Only return memories created at or after this ms-epoch timestamp."
},
"until": {
"type": "integer",
"description": "Only return memories created strictly before this ms-epoch timestamp."
}
}
}
memory_forget
Soft-delete a memory by id. The row is retained for audit but excluded from search/list and removed from the vector index.
Parameters1
id
string
required
UUID of the memory to forget.
Raw schema
{
"type": "object",
"properties": {
"id": {
"type": "string",
"description": "UUID of the memory to forget."
}
},
"required": [
"id"
]
}
memory_supersede
Replace a stale memory with a corrected one. The old row is marked superseded and excluded from future reads; the new row becomes the current version. Use for corrections; use memory_forget for pure deletions.
Parameters5
old_id
string
required
UUID of the memory to replace.
new_content
string
required
New content that supersedes the old memory.
tags
array
optional
written_by
string
optional
session_id
string
optional
Raw schema
{
"type": "object",
"properties": {
"old_id": {
"type": "string",
"description": "UUID of the memory to replace."
},
"new_content": {
"type": "string",
"description": "New content that supersedes the old memory."
},
"tags": {
"type": "array",
"items": {
"type": "string"
}
},
"written_by": {
"type": "string"
},
"session_id": {
"type": "string"
}
},
"required": [
"old_id",
"new_content"
]
}
memory_write_bulk
Write up to 50 memories in a single call. Each entry runs the same path as memory_write (semantic dedup by default; pass force:true per-entry to skip). Embeddings + optimizations run in parallel; D1 inserts are batched. Returns { results: [...] } with one entry per input in the same order — each is { id, created_at, optimized? } on success, { id, created_at, deduped, matched_score } on dedup, or { error } on failure. Free-tier limits apply to the sum: if adding N would exceed 200, the first (200 - existing) succeed and the rest return an error.
Gnosem is a hosted Model Context Protocol server that gives each user one persistent memory store — readable and writeable from every MCP-capable AI client. Vendor-agnostic by design: a fact saved from Claude is visible in ChatGPT, Cursor, Windsurf, Zed, Kimi, and any other MCP client (or plain HTTP tool call) using the same API key.
The problem it solves: every AI vendor has its own proprietary memory. OpenAI's ChatGPT memory doesn't work in Claude. Anthropic's projects don't cross into GPT. Multi-model users retell context in every session. Gnosem is the neutral layer between vendors.
Tools exposed
memory_write — save a fact, preference, decision, or note. content, tags?, written_by? (provenance), session_id?, no_optimize?
memory_search — semantic search across your memories. query, k?, raw?
memory_list — list recent memories in reverse-chronological order. limit?, cursor?, raw?
memory_forget — soft-delete a memory by id
memory_supersede — replace a stale memory with a corrected one
AI-optimized storage
Long memories (>400 chars) are automatically compressed on write to a structured-facts form (Workers AI, llama-3.1-8b-instruct-fast) with a strict prompt: TOPIC=..., PROJECT=..., DECISION=..., STACK=..., PROBLEM=.... The reading LLM ingests the compact form by default — fewer tokens, same meaning. The raw prose is preserved and returned as content_raw; pass raw:true to invert. Pass no_optimize:true on write to skip compression entirely.
Compression is guarded — if the model output isn't actually shorter, gnosem falls back to storing raw only. Fail-open: any AI error still saves the memory.
Quickstart
bash
npx gnosem-install
Auto-detects and configures every MCP-capable client on your machine (Claude Desktop, Claude Code, Cursor, Windsurf, Zed). Prompts for your API key and merges the gnosem entry alongside your existing servers. Restart the affected clients — the five memory tools appear immediately.
Each memory row: id (uuid), content (raw, ≤8000 chars), content_optimized (structured facts, nullable), tags (json array), written_by (provenance — which model wrote it), session_id (opaque grouping), created_at, plus a semantic embedding (BGE-base-en-v1.5, 768-dim) stored in Vectorize.
Every memory tracks provenance so you can see which model contributed which fact. Memories are per-user isolated at both the D1 metadata layer (user_id filter on every query) and the Vectorize layer (metadata index on user_id).
Correction chains via memory_supersede: the old row is marked superseded (excluded from reads) and the new row references it. Soft-delete via memory_forget (excluded from reads and removed from the vector index).
