Memory¶

AEGIS has a persistent, multi-layer memory system. Every conversation builds on what AEGIS already knows about you. The memory data is the crown jewel — it is versioned, human-readable, and always exportable.

Memory layers¶

Layer	What it stores	Examples
Identity	Who you are — always injected into every chat context	Name, role, location, personality notes
Goals	What you're working toward	"Ship AEGIS Phase 12 by end of July"
Projects	Active project contexts	"Project AEGIS: AI-native OS, Phase 12 in progress"
People	Contacts and relationships	"Alice: lead investor, prefers async communication"
Facts	Things AEGIS has learned	"Prefers dark mode. Uses Windows 11."
Conversations	What happened in past sessions	Automatically logged per conversation

The Identity layer is special — those memories are injected into every prompt, even when AEGIS hasn't been asked to recall anything. Everything else is retrieved selectively based on relevance to the current conversation.

Memory Fabric¶

Every memory carries connective metadata (added in v1.1.0) that turns a flat list of facts into a knowledge graph:

Entities — people, places, and things automatically extracted from the memory content at storage time. These form the "mycelial threads" that connect memories to each other through shared entities.

Links — explicit relationships you can draw between memories: - supports — this memory reinforces another - contradicts — this memory conflicts with another (AEGIS will flag the conflict) - elaborates — this memory adds detail to another - supersedes — this memory replaces another (the old one moves to "superseded" lifecycle)

About-time — when the event the memory describes happened, separate from when AEGIS stored it. A memory about a meeting last Tuesday has an about-time of last Tuesday even if AEGIS stored it today.

Lifecycle — the state of the memory: - active — current and in use - superseded — replaced by a newer memory (kept for lineage, not recalled) - expired — past its relevance window - archived — manually moved to cold storage

Recall strength — how often and how recently AEGIS has actually injected this memory into a conversation. High recall = frequently used. AEGIS surfaces this in the Memory panel.

Pin — a pinned memory is never faded, expired, or archived by the system. Use it for facts that must always be available.

Confidence + provenance — where the memory came from affects how much weight AEGIS gives it: - Interview answer → highest confidence - User-stated fact → high confidence - Inferred from chat → lower confidence - Source link — if the memory came from a document, the link is stored

The Interview System¶

The fastest way to build a rich identity layer is through guided interviews. AEGIS has 9 interview sessions covering 46 questions across topics like your background, goals, communication style, preferences, and work context.

Start an interview:

/interview

Choose a session topic, answer the questions in plain language, and AEGIS stores the answers as high-confidence memories.

How AEGIS recalls memories¶

At the start of every chat message:

Identity memories are always included (no search required)
AEGIS searches the remaining memory layers for content relevant to the current message — by semantic similarity, keyword match, entity overlap, and recency
Retrieved memories are injected into the system context before the engine sees your message
The step trace shows exactly which memories were recalled (and their recall strength)

Adding memories manually¶

From the Memory panel: 1. Open Memory → click + Add 2. Enter the content in natural language 3. Set category, tags, and project label 4. Save

From chat:

/memory Remember that Alice prefers email over Slack

Let AEGIS decide: during conversation, AEGIS may ask "Should I remember this?" before storing anything. You can always say yes or no. This is the ask-before-store governance rule.

Editing and deleting¶

Edit: click any memory in the Memory panel to open the inline editor. Change the content, lifecycle, or tags and save.

Delete: click the trash icon. Deletion is permanent — there is no recycle bin. If you want to preserve the memory but stop using it, change its lifecycle to archived instead.

Exporting your memory¶

Memory data is the crown jewel — you can always export it in a human-readable format. Run from the terminal:

python main.py --export-memory

This writes a dated export to data/memory_export_YYYY-MM-DD/ containing: - identity.json — identity layer - goals.json — goals - projects.json — projects - facts.json — all other facts - conversations.json — conversation log - graph.json — Memory Fabric edges (entities, links)

The export format is stable — plain JSON, no proprietary encoding. You own your data.

Project scoping¶

AEGIS supports multiple projects. Conversations and memories are labeled with the active project. Switch projects from the project selector above the chat input.

The project label appears as a 📁 chip on memories in the Memory panel. Filter by project to see only what's relevant to your current work.

Memory and privacy¶

AEGIS is local-first — all memory is stored on your machine in data/memory.db (SQLite). Nothing is sent to external services unless you've explicitly connected a cloud AI engine. The memory store never leaves your machine as part of normal operation.

API keys and secrets are never stored in memory.