Skip to main content

​
remember

Store text in persistent memory with automatic embedding (if configured). Usage: 
remember <text>
Example: 
> remember The greenhouse temperature threshold is 28°C
Remembered (with embedding). (ID: a3f8c2d1, total memories: 43)
Security: Requires memory:write authorization Behavior:
  1. With embedding provider configured:
    • Generates vector embedding via Ollama/OpenAI
    • Stores text + embedding in SQLite (dual-indexed)
    • Enables semantic search via recall
  2. Without embedding provider:
    • Stores text with FTS5 keyword indexing only
    • Graceful fallback (zero config required)
Example (no embedding): 
> remember Sensor A1 calibrated on 2026-03-01
Remembered. (ID: b7e4f931, total memories: 44)
Storage:
  • Backend: SQLite (default) or NoopMemory
  • Auto-generates UUID for each entry
  • Timestamps with creation time
  • Metadata support (tags, source, etc.)

​
recall

Search memory using hybrid FTS5 + vector similarity with RRF fusion. Usage: 
recall <query>
Example (hybrid search): 
> recall greenhouse temperature
Found 3 memories:
  1. [score:0.87] [2026-03-02 14:15] The greenhouse temperature threshold is 28°C
  2. [score:0.72] [2026-03-01 09:30] Temperature sensor installed in greenhouse zone B
  3. [score:0.68] [2026-02-28 16:45] Automated cooling activates at 27°C
Example (FTS-only fallback): 
> recall sensor calibration
Found 2 memories:
  1. [2026-03-01 10:20] Sensor A1 calibrated on 2026-03-01
  2. [2026-02-25 14:00] Annual sensor calibration schedule: March 1st
Security: Requires memory:read authorization Search Modes:

​
Hybrid Search (FTS5 + Vector + RRF)

When embedding provider is configured:
  1. FTS5 keyword search: Traditional full-text search on indexed tokens
  2. Vector similarity: Cosine similarity between query embedding and stored embeddings
  3. RRF fusion: Reciprocal Rank Fusion combines both rankings
Benefits:
  • Finds semantically similar content even with different wording
  • Balances keyword precision with semantic recall
  • Robust to synonyms and paraphrasing
Example: 
Query: "heat management"
Matches: "temperature threshold", "cooling system", "thermal control"
(no exact keyword overlap, matched via semantic similarity)
When no embedding provider configured:
  • Pure SQLite FTS5 keyword matching
  • Fast, lightweight, zero external dependencies
  • Graceful degradation (no errors)
Limits:
  • Default: 5 results
  • Sorted by relevance (hybrid score or FTS rank)
No Results: 
> recall nonexistent query
No memories found.

​
Memory Search in LLM Pipeline

When you send queries to the LLM (via ask or free-form text), OneClaw automatically:
  1. Searches memory for relevant context (top 5 results)
  2. Injects matches into LLM prompt as “Related data from memory”
  3. LLM incorporates stored knowledge into response
Example: 
> ask What's the temperature limit for the greenhouse?
Internal prompt sent to LLM: 
Related data from memory:
- [02/03 14:15] The greenhouse temperature threshold is 28°C
- [01/03 09:30] Automated cooling activates at 27°C

User question: What's the temperature limit for the greenhouse?
LLM response: 
The greenhouse temperature threshold is 28°C. The automated cooling system 
activates slightly earlier at 27°C to prevent overheating.
Offline Mode: If no LLM provider is configured, memory search still works: 
> ask greenhouse temperature
[Offline mode] No LLM provider configured.
2 related entries found in memory.

​
Embedding Configuration

Enable vector search in config/oneclaw.toml: 
[embedding]
provider = "ollama"  # or "openai"
model = "nomic-embed-text"  # 768 dimensions
Ollama (local, offline-capable): 
[embedding]
provider = "ollama"
model = "nomic-embed-text"  # 768d
url = "http://localhost:11434"
OpenAI (cloud): 
[embedding]
provider = "openai"
model = "text-embedding-3-small"  # 1536d
api_key = "sk-..."  # or set OPENAI_API_KEY env var
Verify embedding status: 
> status
...
Embedding: nomic-embed-text (768d) — 38 embedded / 42 total
...

​
Memory Backend

Default: SQLite with FTS5 + vector extension Storage location:
  • workspace/oneclaw.db (SQLite file)
  • Auto-created on first remember
Schema:
  • memories table: id, content, created_at, metadata (JSON)
  • memories_fts FTS5 virtual table: keyword index
  • memory_embeddings table: id, embedding (BLOB), dimensions
Persistence:
  • Survives runtime restarts
  • No data loss on clean shutdown

Build docs developers (and LLMs) love

Get started for freeTalk to us