Memory

🧠 ATOM Memory Guide

ATOM's memory engine gives your assistant real recall: it can remember who you are, what you've uploaded, and what you've asked before — all running locally.

⚡ New in April 2025 — Smarter Memory Pipeline

ATOM now includes an advanced memory pipeline for sharper, more relevant context injection and reasoning:

✅ Semantic Deduplication

Filters out near-duplicate memory chunks using cosine similarity to reduce repetition and noise.

✅ LLM-Powered Reranking

Uses your selected local model (like mistral or cogito) to re-rank vector memories for relevance before they’re injected into the prompt.

✅ Thread Awareness

Tracks the last 5 user messages and includes them as [Recent Thread] in the prompt to maintain coherent, flowing conversations.

✅ Cogito Prompt Format

When using the cogito model, ATOM switches to a structured prompt format to support logical reasoning:

# System:
<persona>

# Memory:
<context>

# Question:
<your input>

# Thoughtful Answer:

These upgrades make ATOM far more consistent and reflective across multi-turn sessions.

💾 Memory Types

ATOM supports multiple memory layers, each with a different lifespan and retrieval method:

🔹 Identity Memory

Stores things like your name, preferences, tone, and personality traits.

  • Injected often
  • Never expires unless cleared

🔹 Chat Memory

Captures past conversation history, prioritized by relevance and recency.

  • Retrieved using vector search
  • Pruned automatically over time

🔹 File Memory

Stores content from uploaded files, chunked into searchable segments.

  • Searchable by question
  • Context-aware during file Q&A

🔹 Summary Memory

Reflections and rollups that summarize what the assistant has learned.

  • Triggered automatically every 10 turns
  • Includes file summaries and self-generated thoughts

📥 File Upload Pipeline

When you upload a file, ATOM:

  1. Parses and chunks the text
  2. Scores each chunk (skips junk)
  3. Runs summarization
  4. Stores content in vector memory

Supported formats:

  • .txt, .pdf, .docx, .zip
  • Scanned PDFs get OCR'd automatically

🔁 Reflection Cycle

Every 10 messages, ATOM:

  • Generates reflections for each memory type
  • Purges low-score or expired chunks
  • Reprioritizes recent and useful data

These reflections are visible in the Memory Dashboard.

🧪 Advanced Behaviors

  • Score thresholds prevent noisy memory from being used
  • Chat context is rebuilt fresh each turn
  • Active file context is injected alongside vector matches
  • Streaming is disabled automatically for Discord-based messages to prevent partial outputs

Oh yeah..... it has a built in discord bridge, add it as a legit bot to your server and talk to any of your ollama models ON DISCORD.

ATOM’s memory engine is built for local, smart, and modular reasoning. Most of the time. YMMV.

Want to see how tools work under the hood?
→ Head to the Tools page next.

InstallTools