Search & scope - Memory Crystal

Search tools help the AI find the right memory without looking in the wrong places. All searches respect channel, client, and knowledge-base isolation boundaries.

What this means in practice

This includes:

semantic search over recent messages (STM)
scoped message visibility by channel
time-range filtering for recent continuity
knowledge-base-aware query behavior
client/channel isolation rules (enforced server-side)

crystal_search_messages — Semantic message search

Search short-term memory (recent messages) by semantic similarity. Complements crystal_recall (which searches long-term memories).

Parameters

Parameter	Type	Required	Description
`query`	string	Yes	The search query. Embedded and matched against message content.
`limit`	number	No	Maximum results to return. Default: 10. Range: 1–20.
`channel`	string	No	Scope search to a specific channel. If not provided, searches all accessible channels.
`sinceMs`	number	No	Only return messages after this Unix timestamp (in milliseconds). Useful for time-range queries.

Returns

{
  “query”: “authentication strategy”,
  “results”: [
    {
      “_id”: “msg_abc123”,
      “role”: “user”,
      “content”: “How should we implement auth in the new API?”,
      “channel”: “engineering”,
      “timestamp”: 1713314400000,
      “score”: 0.89
    },
    {
      “_id”: “msg_def456”,
      “role”: “assistant”,
      “content”: “I recommend using OAuth2 with JWT tokens for stateless auth.”,
      “channel”: “engineering”,
      “timestamp”: 1713314460000,
      “score”: 0.85
    }
  ],
  “limit”: 10,
  “channel”: “engineering”,
  “sinceMs”: null
}

Examples

Basic message search:

{
  “query”: “what did we decide about the database?”
}

Search within a specific channel:

{
  “query”: “deploy process”,
  “channel”: “devops”
}

Search recent messages only (last 7 days):

{
  “query”: “API authentication”,
  “sinceMs”: 1712104800000,
  “limit”: 5
}

Search engineering channel since a specific time:

{
  “query”: “performance optimization”,
  “channel”: “engineering”,
  “sinceMs”: 1713000000000,
  “limit”: 15
}

Channel filtering

What channels do

Isolate conversations — messages in one channel are not visible to queries scoped to another channel
Enable cross-client scenarios — the same user can have separate channels for different integrations (e.g., Slack vs Telegram)
Preserve context boundaries — memory recall respects channel scope by default

Using channels

If you omit channel, searches return results from all channels the user has access to. To narrow results:

Provide the channel name — Pass the channel identifier (e.g., ”slack”, ”telegram”, ”claude-code”)
Verify visibility — Server-side rules enforce that you only see messages within accessible channels
Combine with sinceMs — Time range + channel gives fine-grained control

Example: Multi-channel scenario

A user interacts with Memory Crystal via both Slack and Telegram. Each has its own channel. To search only Slack messages:

{
  “query”: “project timeline”,
  “channel”: “slack”
}

To search both, omit the channel parameter (or make two calls with different channels).

Time-range filtering with sinceMs

Use sinceMs to retrieve messages after a specific Unix timestamp (milliseconds):

// Messages from the last 7 days
const sevenDaysAgo = Date.now() - (7 * 24 * 60 * 60 * 1000);

// Messages from a specific date (2026-04-16)
const specificDate = new Date(“2026-04-16”).getTime();

Examples

Messages from today onwards:

{
  “query”: “decisions made”,
  “sinceMs”: 1713004800000
}

Messages from a specific week:

{
  “query”: “deployment issues”,
  “sinceMs”: 1712000000000,
  “channel”: “devops”
}

When to use vs alternatives

crystal_search_messages (recent, exact context)

Use when: You need to find recent messages by semantic meaning, preserving context and timestamps. Best for: Session continuity, recent decision tracking, finding exact phrasing from recent conversations. Example: “What did we just discuss about authentication?”

crystal_recall (long-term, high-confidence memories)

Use when: You want to find durable, extracted facts and decisions. Better for understanding patterns and history. Best for: Getting the gist of a topic without reading full conversation history. Example: “What have we decided about authentication?”

Decision guide

Scenario	Tool
Find an exact recent message or phrasing	`crystal_search_messages`
Get the summary/decision from a topic	`crystal_recall`
Get context on why a decision was made	`crystal_why_did_we`
Get all information about a topic	`crystal_what_do_i_know`
Find the most recent N messages chronologically	`crystal_recent`
Query a knowledge base	`crystal_query_knowledge_base`

Scoping and isolation

Memory Crystal enforces isolation at multiple levels:

User-level isolation — Users never see each other’s messages or memories (enforced by userId)
Channel-level isolation — Messages in one channel don’t appear in queries scoped to another (enforced by channel filtering)
Knowledge-base isolation — KB queries only return chunks from the specified knowledge base
Age-based visibility — Archived memories and expired STM messages are hidden by default

Server-side guarantee: All filtering is enforced server-side. Client code cannot bypass isolation rules.

Error handling

Error	Cause	Resolution
query is required	No query parameter was provided.	Provide a non-empty search string.
Embedding service unavailable	The embedding backend is unreachable.	Retry; check service health.
Memory Crystal recall degraded	Same as above.	Retry later.
Invalid channel	Channel identifier is malformed or inaccessible.	Verify the channel name exists.
Invalid sinceMs	sinceMs is not a valid Unix timestamp.	Use `Date.now()` or a numeric millisecond timestamp.
Empty results	No messages matched the query/filters.	Try a broader query or remove the channel/time filter.

How it actually works

Key repo surfaces:

mcp-server/src/tools/search-messages.ts — tool definition and HTTP handler
convex/crystal/messages.ts — message storage and search logic
convex/crystal/mcp.ts — API endpoint routing
convex/crystal/recall.ts — scoping and filtering rules

The flow:

Tool validates query, limit, channel, and sinceMs parameters
Query is embedded using OpenAI’s text-embedding-3-small
ConvexClient POSTs to /api/mcp/search-messages with embedding
Backend applies channel and time filters
Vector search retrieves top-K matching messages
Results are returned with scores and timestamps

Channel and user isolation is enforced at the database level, not in the tool code.

Common mistakes

Treating search and recall as if they are identical (they’re designed for different purposes)
Forgetting that scoped channels affect results (all searches are channel-aware)
Not using sinceMs for time-range queries (just relying on default recency)
Assuming cross-channel visibility when channels are set (isolation is enforced)
Using crystal_search_messages for long-term patterns (use crystal_recall instead)

Source of truth

Primary files behind this page:

mcp-server/src/tools/search-messages.ts — tool definition
convex/crystal/messages.ts — backend logic and filtering
convex/crystal/recall.ts — scoping rules
convex/crystal/knowledgeBases.ts — KB visibility logic
convex/crystal/mcp.ts — HTTP endpoint routing

​What this means in practice

​crystal_search_messages — Semantic message search

​Parameters

​Returns

​Examples

​Channel filtering

​What channels do

​Using channels

​Example: Multi-channel scenario

​Time-range filtering with sinceMs

​Examples

​When to use vs alternatives

​crystal_search_messages (recent, exact context)

​crystal_recall (long-term, high-confidence memories)

​Decision guide

​Scoping and isolation

​Error handling

​How it actually works

​Common mistakes

​Source of truth

What this means in practice

crystal_search_messages — Semantic message search

Parameters

Returns

Examples

Channel filtering

What channels do

Using channels

Example: Multi-channel scenario

Time-range filtering with sinceMs

Examples

When to use vs alternatives

crystal_search_messages (recent, exact context)

crystal_recall (long-term, high-confidence memories)

Decision guide

Scoping and isolation

Error handling

How it actually works

Common mistakes

Source of truth