Query Memory & Context Synthesis¶

kompend includes a lightweight, native memory layer that allows the system to learn from past interactions. This feature enables long-term context retention, ensuring that the wiki evolves to better match your specific needs, terminology, and preferences over time.

Overview¶

The memory layer operates on a dual-storage model for optimal performance and flexibility:

Semantic Search (zvec): Past queries and answers are embedded and stored in .kompend/memory.db, allowing the system to retrieve highly relevant past interactions based on semantic similarity.
Chronological Log: An append-only log in the memory_log table of .kompend/kompend.db maintains a strict timeline of interactions, which is essential for periodic context synthesis.

CLI Usage¶

1. Saving Queries During Search¶

When using kompend query, you can append the --remember flag. If the query is successfully synthesized by the LLM, both the query and the generated answer will be automatically saved to the memory store.

kompend query "What are the acoustic intelligibility requirements?" --remember

2. Manual Memory Management¶

You can interact with the memory store directly using the kompend memory command group.

Add an entry manually:

kompend memory add "How do I configure Ollama?" "Set the LLM model to ollama/..." --feedback "Prefer concise, step-by-step answers"

Search past interactions:

kompend memory search "Ollama configuration" --top-k 3

Synthesize context: Periodically, you can instruct kompend to review recent memory entries and distill them into a living system context file. This file (sources/_memory/system_context.md) is automatically picked up during compilation and querying to guide the LLM's behavior.

kompend memory synthesize --limit 20

Note: In the current implementation, synthesis generates a thematic summary of recent queries. Future iterations will use an LLM to generate deeper behavioral insights (e.g., "The user prefers bullet points and focuses heavily on acoustic metrics").

MCP Server Integration¶

For AI agents, the memory layer is fully exposed via the kompend MCP server, allowing autonomous agents to build and leverage long-term context without shell access.

Available MCP tools:

memory_add: Save a query/answer/feedback triple.
memory_search: Retrieve semantically similar past interactions to inform the current agent context.
memory_synthesize: Trigger the periodic distillation of recent interactions into the system_context.md file.

Example Agent Workflow¶

Agent answers a user query about the wiki.
Agent calls memory_add with the query, its answer, and any implicit/explicit user feedback.
On subsequent, related queries, the agent calls memory_search to see how similar questions were handled previously.
Periodically, the agent (or a background cron job) calls memory_synthesize to update the global system context.

Enhanced RAG Pipeline¶

kompend includes optional RAG enhancement techniques that improve retrieval quality. Each is a plug-in step configurable via kompend.toml under [search]:

Feature	Config	Default	Description
Multi-Query Expansion	`search.expansion.enabled`	`false`	N query variants + RRF merge (RAG-Fusion)
CRAG	`search.crag.enabled`	`false`	LLM relevance grading, drop noise
Neural Re-ranker	`search.reranker.enabled`	`false`	Cross-encoder post-fusion ranking
HyDE	`search.hyde.enabled`	`false`	Hypothetical answer embeddings
Contextual Retrieval	`search.contextual.enabled`	`false`	Situating context for chunk embeddings

All features default to disabled. Enable only what you need. See RAG Pipeline for full documentation.

Configuration¶

The memory layer respects the global wiki directory context. By default, it looks for the .kompend/ directory in the current working directory (or the directory specified by --wiki-dir in the CLI).

No additional configuration is required to enable the memory layer, though ensuring your compile.embedding.model is configured correctly will provide the best semantic search results.