Skip to content

memory.md

Latest commit

 

History

History
130 lines (86 loc) · 3.75 KB

memory.md

File metadata and controls

130 lines (86 loc) · 3.75 KB

Memory System

Agentis has a built-in persistent memory system backed by IndexedDB — the browser's native database. No external service, no PocketBase, no server required.

How It Works

Every time an agent completes a task, a memory entry is automatically saved:

  • Key — a normalized version of the task description
  • Value — the first 400 characters of the output
  • Source — auto (agent-saved) or manual (you added it)

On future tasks, agents retrieve relevant memories and include them as context in their system prompt.

Storage

Old system (localStorage) New system (IndexedDB)
Capacity ~5 MB Gigabytes
Structure Flat JSON array Indexed database
Survives browser clear No Yes
Query speed Linear scan Indexed lookup
Server required No No

Memory Entry Fields

Each memory entry stores:

Field Description
agentId Which agent persona this belongs to
key Short identifier (normalized task text)
value Stored content (up to 400 chars)
source auto or manual
category episodic, semantic, procedural, or general (auto-inferred)
tags Optional string array
importance 0–1 score, decays over time
accessCount How many times this memory was read
lastAccessed Timestamp of last read

Categories (Auto-Inferred)

When a memory is saved, Agentis automatically classifies it:

Category Triggered by keywords
Procedural code, function, implement, build, deploy, debug, test, api
Semantic fact, concept, definition, knowledge, theory, explain
Episodic session, task, completed, result, output, previously
General everything else

Importance & Decay

Every memory starts with an importance score:

  • Manual memory → 0.8
  • Auto-saved → 0.5

Each time the memory is accessed (read by an agent): importance increases by +0.1 (capped at 1.0).

Over time, importance decays exponentially:

importance = stored_importance × e^(−decay_rate × days_since_last_access)

Default decay rate: 5% per day. A memory at 0.5 importance drops to ~0.3 after 10 days of no use.

You can change the decay rate under Settings → Memory using the slider.

Managing Memory

View & Search

Go to Sessions → Memory tab or Settings → Memory.

  • Filter by agent persona
  • Search across keys, values, and tags
  • See importance score bar and category badge for each entry

Add Manually

In Sessions → Memory:

  1. Select the persona
  2. Click + Add memory
  3. Enter a key and value

Export

  1. Go to Settings → Memory
  2. Click Export
  3. A .json file downloads with all entries

Import

  1. Go to Settings → Memory
  2. Click Import
  3. Select a previously exported .json
  4. Duplicate entries are automatically skipped — only new entries are added

This is how you can share or sync memory between machines (export on one, import on another).

Prune Decayed

Click Prune Decayed in Settings → Memory to delete all entries whose importance has fallen below 5%. Keeps your memory store lean and relevant.

Migration from Old System

If you were using an older version of Agentis that stored memories in localStorage, they are automatically migrated to IndexedDB on first load. Nothing is lost.

Coming Soon

  • Semantic search — find memories by meaning, not just keyword matching
  • Memory consolidation — automatically summarize groups of related memories into one
  • Cross-device sync — optional sync token to share memory between browsers without a server
  • Memory timelines — browse what each agent has learned over time with a visual timeline