Manage Memory

Snippbot agents have persistent memory across sessions. Every conversation is automatically captured as an episode, entities are extracted into a knowledge graph, and recall improves over time. This guide explains how to search, explore, configure, and manage memory through the UI.

Expand

Accessing the Memory page

Navigate to the Memory page using the sidebar. The page has five tabs:

Tab	Component	What it shows
Explore	MemoryNebula	A 2D/3D particle visualization of memory clusters. Click clusters to view related episodes. Zoom and pan to navigate.
Timeline	EpisodeStream	A chronological list of all memory episodes, grouped by date (Today, Yesterday, etc.) with infinite scroll.
Search	SearchInterface	Search across all memory with three modes, filters, and highlighted results.
Knowledge Graph	Knowledge Graph	A force-directed entity-relation visualization with color-coded nodes and interactive exploration.
Stats	MemoryStats	Aggregate memory metrics — total episodes, entities, relations, importance distribution, and recall feedback statistics.

The Memory page header displays summary statistics: episode count, vector count, graph node count, and storage size.

Searching memory

The Search tab provides three search modes, each suited to different types of queries.

1. Navigate to the Memory page and select the Search tab.

2. Enter your query in the search box. The search is debounced (300ms) so results appear as you type.

3. Select the search mode:

   Mode	Best for	How it works
   Hybrid (default)	General queries	Combines keyword and semantic search using Reciprocal Rank Fusion. A query analyzer auto-detects the optimal blend between keyword and vector weights.
   Semantic	Conceptual questions, “how do I…” queries	Uses 384-dimensional vector embeddings to find episodes by meaning, not just exact words. Finds results even without keyword overlap.
   Keyword	Error codes, version numbers, exact terms	Uses FTS5/BM25 full-text search. Supports phrase matching, boolean operators, and prefix wildcards. Includes a 7-day recency boost.

4. Review results — each result shows the matched episode with highlighted text, a relevance score, and the source type.

Tip

Use Keyword mode when searching for specific error messages, package names, or technical identifiers. Use Semantic mode when you remember the concept but not the exact words. Use Hybrid for everything else.

Exploring the timeline

The Timeline tab (EpisodeStream) shows all memory episodes in chronological order.

1. Navigate to the Memory page and select the Timeline tab.

2. Scroll through episodes — they are grouped by date (Today, Yesterday, This Week, etc.) and load more as you scroll down (infinite pagination).

3. Expand an episode by clicking on it to see the full content, importance score, valence, source type, and linked entities.

4. Delete an episode by expanding it and clicking Delete. This removes the episode from episodic memory, the vector index, and any entity links.

Viewing the knowledge graph

The Knowledge Graph tab shows a force-directed visualization of all entities and their relationships.

1. Navigate to the Memory page and select the Knowledge Graph tab.

2. Explore the graph — nodes are color-coded by entity type (languages, frameworks, tools, concepts, people). Drag nodes to rearrange the layout.

3. Click on a node to see its properties: name, type, importance score, and connected relationships.

4. Filter by entity type using the filter controls to focus on specific categories (e.g., show only frameworks and tools).

The graph is built automatically from your conversations. As you interact with your agent, entities and relations are extracted and added to the graph. The graph view is read-only — nodes and edges are managed by the memory system as conversations progress.

Exploring the nebula

The Explore tab (MemoryNebula) provides a visual cluster map of your agent’s semantic memory.

1. Navigate to the Memory page and select the Explore tab.

2. Navigate the visualization — zoom in and out, and pan across the canvas to explore memory clusters.

3. Click on a cluster to view the episodes it contains. Clusters represent groups of semantically similar memories.

Viewing memory stats

The Stats tab shows aggregate metrics for the selected agent’s memory.

1. Navigate to the Memory page and select the Stats tab.

2. Review the overview — total episodes, total entities, total relations, average importance, and average valence.

3. Check the importance distribution — a histogram showing how episodes are distributed across importance levels.

4. Review recall feedback — total recalls, used count, ignored count, and usage rate. This shows how often recalled memories actually influenced the agent’s responses.

How auto-capture works

Every conversation is automatically captured as a memory episode. You do not need to do anything to enable this — it happens in the background after every chat response.

The capture process:

1. The user message and assistant response are combined into a single content block.
2. The Sensory Buffer filters out trivial exchanges (under 100 characters) and near-duplicates (over 80% word overlap).
3. Content is normalized (Unicode cleanup, whitespace normalization, truncation at 10,000 characters).
4. Importance is scored based on content keywords (“critical”, “decision”, “breakthrough”), content length, and source type. Failures get a boost (+0.2) because agents learn from mistakes.
5. Valence (sentiment) is scored from -1.0 to +1.0 based on positive and negative keyword counts.
6. The episode is stored, full-text indexed, embedded as a 384-dimensional vector, and entities are extracted into the knowledge graph.

Note

The meta-cognitive layer continuously adjusts episode importance based on whether recalled memories actually influence the agent’s responses. Useful memories get boosted (+0.05); ignored memories get decayed (-0.02). An Ebbinghaus forgetting curve applies time-based decay weighted by importance.

Configuring memory settings

1. Navigate to Settings in the sidebar.

2. Select the Memory section.

3. Adjust the following settings:

   Setting	Default	Description
   Memory Enabled	On	Master toggle for the entire memory system. When off, no episodes are captured or recalled.
   Auto Recall	On	Automatically surface relevant memories during conversations. When off, memory exists but is not injected into chat context.
   Recall Scope	All	Which memories to search during recall: All (global across all conversations), Project (only memories from the current project), Session (current session only), or None (disable recall without disabling capture).
   Retention Policy	Forever	How long to keep episodes: Forever, 1 Year, 6 Months, 3 Months, or 1 Month. Episodes older than the retention period are automatically pruned.
   Max Episodes	10,000	Maximum number of stored episodes (range: 100 to 100,000). When the limit is reached, the oldest low-importance episodes are removed first.
   Auto Summarize	On	Automatically compress older episodes via LLM summarization to save storage and improve retrieval quality.
   Summarize Threshold	30	Number of messages before auto-summarization triggers (range: 7 to 90 days).
   Min Relevance Threshold	0.25	Minimum similarity score (0.0 to 1.0) for injecting a memory into chat context. Below this threshold, nothing is injected — silence is better than noise.

4. Click Save to apply changes.

Tip

If your agent is surfacing too many irrelevant memories, increase the Min Relevance Threshold from 0.25 to 0.4 or higher. If it is missing important context, lower it toward 0.1.

Configuring privacy settings

Privacy settings are located in the same Settings page, under the Memory section.

1. Navigate to Settings in the sidebar and select the Memory section.

2. Adjust the following privacy settings:

   Setting	Default	Description
   Redact Secrets	On	Automatically strips API keys, passwords, tokens, and other secrets from episodes before they are written to the database.
   Anonymize Names	Off	Removes or replaces personal names in stored episodes to protect identity information.
   Local Only	On	Enforces that all memory data remains on the local machine. No cloud sync, no data transmission.
   Exclude Patterns	Empty	Regex or glob patterns to exclude from memory storage. Use this to prevent specific content from being captured (e.g., *.env, sensitive file paths, internal URLs).

3. Click Save to apply changes.

Note

All memory data — episodes, embeddings, knowledge graph, and vector indices — is stored in SQLite databases on your machine. The Local Only setting is on by default and ensures nothing is synced, uploaded, or transmitted. Your agent’s memory exists only on your hardware.

Exporting and clearing memory

Export memory data

1. Navigate to the Memory page.

2. Click Export Memory in the page actions.

3. A JSON file is downloaded containing all episodes and their metadata (content, importance, valence, timestamps, entities, and relations).

Clear all memory

Caution

Clearing all memory is irreversible. The agent loses all episodic knowledge, the vector index, and the knowledge graph from all past sessions.

1. Navigate to the Memory page.

2. Click Clear All Memory.

3. Confirm the action in the confirmation dialog. This performs a cascading delete of all episodes, the vector index, and the knowledge graph.

Delete individual episodes

1. Navigate to the Memory page and select the Timeline tab.

2. Click on an episode to expand it.

3. Click Delete to remove that specific episode from memory.

Automatic pruning

Episodes are automatically pruned based on two settings:

• Retention Policy — episodes older than the configured period are removed on a rolling basis.
• Max Episodes — when the episode count exceeds the limit, the oldest low-importance episodes are removed first.

Tips for effective memory

• Review what your agent knows — use Semantic search to check what the agent remembers before starting important work.
• Tune the relevance threshold — increase it if the agent injects too much noise; decrease it if it misses useful context.
• Use Project scope — for project-specific work, set Recall Scope to Project so the agent only recalls relevant project memories.
• Keep episode counts reasonable — very large episode stores (above 10,000) can slow retrieval. Use retention policies to prune old episodes.
• Monitor the knowledge graph — check which entities and relations the agent has learned. This helps you understand how the agent connects concepts.
• Clear and restart if an agent has accumulated contradictory memories — sometimes a fresh start gives better results.

Related

• Memory concepts — the 5-tier architecture, hybrid search, knowledge graph, and privacy model explained in depth