Your AI team's long-term memory — structured, alive, and always connected.
Memoria goes beyond a document store. It's a living knowledge graph where every vault note, KB document, and agent insight lives as a node — connected by semantic edges, queryable by your entire AI team, and intelligently maintained through confidence scoring and decay.
Knowledge Base vs Memoria — at a glance
| Capability | Knowledge Base | Memoria |
|---|---|---|
| What it stores | Documents — PDF, DOCX, Drive | Structured nodes — facts, insights, entities |
| How agents use it | Semantic search (RAG) | Graph query + semantic edges |
| Auto-populated | ❌ Requires upload | ✅ Crystallisation from every conversation |
| Confidence over time | — | ✅ Decays & reinforces |
| Obsidian / vault import | ❌ | ✅ ZIP/Markdown with YAML frontmatter |
| Graph visualisation | ❌ | ✅ Force-directed D3 graph |
| Health monitoring | ❌ | ✅ Orphans, duplicates, stale, low-confidence |
How Memoria works — step by step
Before ingesting anything, understand how AEGIS OS organises knowledge. The Knowledge Base is a flat document store — you upload files, they get chunked and indexed for semantic search. Memoria is a graph — nodes (facts, entities, decisions, insights) are connected by typed edges (related_to, crystallized_from, contradicts, references). The same fact can live in both layers and serve different retrieval paths.
16 documents
142 chunks indexed
→ RAG retrieval
5,203 nodes · 599 edges
→ Graph query
Go to Memoria Explorer (sidebar → 🧠 Memoria). Drag a ZIP of your Obsidian vault, or drop a single .md file.
AEGIS reads YAML frontmatter — tags, aliases, links — without calling any LLM. Your pre-tagged notes are ingested in seconds.
For untagged notes, AEGIS runs LLM enrichment as a background job, batching up to 5 notes per second.
Tip: If you use Obsidian with the Dataview or Templater plugins, run your local pipeline first to pre-populate tags: in every note's frontmatter. The ingest will be instant and zero LLM calls needed.
Once ingested, the Graph Explorer renders a live force-directed D3.js graph. Nodes are colour-coded by type: ● blue = vault notes, ● green = KB documents, ● orange = agent insights. Node opacity reflects confidence — fully opaque = high confidence, faded = decaying. Superseded nodes are outlined in red.
Click any node to open the detail panel — see its full content, tags, confidence score bar, linked nodes, and an option to approve, reinforce, or supersede it. Use the type filter chips and search bar to focus the graph on what you need.
Every time an agent responds to a conversation, Crystallisation runs in the background. It extracts 2–3 key insights from the response, saves them as orange agent_insight nodes, and creates semantic edges linking them to related existing nodes. No manual work required. The graph grows smarter with every conversation.
Boardroom sessions are always crystallised — even if you've hit the per-tenant rate limit for regular chats. High-quality crystallisation requires responses of at least 350 characters and excludes short acknowledgements ("OK, got it", "Sure").
Every agent_insight node starts at 0.6 confidence. Human-authored nodes start at 1.0 and never decay. Agent insights decay approximately 5% per month — a fact not confirmed in 6 months drops to ~0.75, and in 18 months to ~0.6. When you or another source confirms the same fact, the node is reinforced (+0.15, capped at 1.0) and the clock resets.
If a fact is replaced by a newer version, you can supersede it — the old node's confidence drops to 0.1 and is linked to the new one. This preserves history while making clear which version is current.
Click Run Health Check in the Memoria Explorer sidebar. AEGIS scans the full graph for six classes of issues and returns a health score from 0–100. Aim for 80 or above. A fresh vault import will typically score 70–80; after a week of crystallisation and a few enrichment passes it will reach 85+.
When to use Memoria
Frequently asked questions
tags:, aliases:, and wikilinks from YAML frontmatter without calling any LLM — your pre-processed notes ingest in seconds. For notes without frontmatter, a background enrichment job will tag them using your LLM configuration.agent_proposed) decay approximately 5% per month using an exponential formula: score × 0.9983^(days_since_reinforced). The floor is 0.1 — a node never reaches zero. Human-authored nodes (uploaded by you) start at 1.0 and do not decay. When a fact is confirmed by another source, reinforce the node to add 0.15 (capped at 1.0) and reset the decay clock.Upload a vault, start talking to your agents,
and watch Memoria build itself — one conversation at a time.