Contact usTry Agent Tale
Skip to content

DocumentationLore

Lore

Lore (/lore, its own LORE sidebar group) gives agents persistent memory and specialized knowledge (v3.76+). Agents search uploaded documents (PDF, Markdown, code…), accumulate experience from chats and Legend runs, and store it in layered memory. Relevant context is added automatically to chats and Legend nodes.

flowchart LR
    subgraph Inputs
      D[Documents: PDF, code, text]
      S[Agent chats, Assistant and Legend runs]
    end
    subgraph Lore
      C[Documents: Knowledge base]
      M[Memory: Layers L0-L3]
      G[Map: Entity graph]
    end
    subgraph Outputs
      A[Specialized agent with memory]
      T[Toolset / Legend lore node]
    end
    D --> C
    S --> M
    C --> G
    M --> A
    C --> A
    C --> T

Three core capabilities

Tab Type Behavior
Documents corpus Uploads are chunked, embedded and served through hybrid search.
Memory memory Chats and Legend executions are distilled into persistent L0 → L1 → L2 → L3 experience. It fills automatically, not manually.
Map Graph Extracts people, places, concepts and relationships into an interactive graph.

On narrow screens, the list appears first; selecting an item opens detail with a back button.

1. Documents: Knowledge base

  1. Choose New Lore base and give it a name.
  2. Select it, then Upload file or Paste text with a title and content.
  3. Track Queued → Processing… → Ready, or Couldn't process. After two minutes, a “Processing for Xm” hint appears. Ready documents show their section count.

  4. Try Search this Lore base.

  5. Formats: .txt, .md/.markdown, .csv, .json, .pdf, plus code/config files (.go, .py, .js, .ts, .rs, .java, .rb, .sh, .yaml/.yml, .toml). Maximum 8 MB per file. Scanned PDFs without text cannot be processed.

  6. Hybrid search: dense cosine similarity and BM25 lexical results are combined with Reciprocal Rank Fusion. Enabling Map adds the entity graph as a third signal.
  7. URL import: available via POST /api/v1/lore/bases/:slug/documents/url with private-network protection; the web UI has no URL field yet.

2. Layered memory (L0 → L3)

Four distillation layers model short- and long-term memory:

flowchart TD
    L0[L0: Raw transcripts of chats and Legend steps]
    L1[L1: Facts and preferences]
    L2[L2: Session-level scenarios]
    L3[L3: Persistent core memory]
    L0 -->|at least 4 new turns| L1
    L1 -->|every 6 atoms| L2
    L2 -->|every 3 scenarios| L3
  1. L0: automatically captures raw messages from agent chat, Assistant and Legend.
  2. L1: after at least four new turns, discards repetitions and trivial exchanges and extracts facts such as “prefers Go to Python”.
  3. L2: updates a session summary every six new atoms.
  4. L3: updates the high-level understanding of you and your project every three new scenarios.

The Memory tab shows a separate base per agent, with CONVERSATIONS (sessions, turn counts, expandable transcript), DISTILLED INSIGHTS (Fact / Scenario / Core cards) and search.

Summary quality depends on configuration

Real LLM summaries require an administrator-configured distillation provider under Settings → Developer → AI Configuration. Otherwise, heuristic distillation still extracts L1 atoms; L2/L3 contain compilations labeled as not summarized.

3. Map: Entity graph

Enable Show on Map (On Map) in Documents to extract people, places, objects, groups, events, concepts and relationships from new content. Alternate names are merged. Map lists only enabled bases and renders their graphs interactively.

4. Connect Lore to agents and workflows

Surface How
Agent chat Conversations are written to memory; the agent's memory base binds automatically. Relevant context precedes the system prompt in later conversations with your own agent.
Toolset Settings → Developer → MCP Bridge → Lore imports a base as search_lore and list_documents tools. See Toolsets.
Legend A lore node searches a base using Slug / Query / Top-K. See Legend.
Assistant search_lore searches your bases. See Assistant.

Direct document-base bindings currently use /api/v1/lore/agents/:id/bindings.

5. Portable memory: .lorepack

  • Export: GET /api/v1/lore/bases/:slug/export creates an archive of documents, sections and vectors.
  • Import: POST /api/v1/lore/bases/import loads it. Matching embedding models reuse vectors; different models re-embed the content.

These actions are API-only; the Lore screen has no import/export buttons yet.

Security and privacy

  • Each base belongs to its creator and is private by default.
  • URL imports block private addresses such as 10.x, 192.168.x and localhost.
  • Upload, search and memory distillation currently cost no credits. A Legend lore node costs 1 ⚡; see Credits.

Related