An AI agent with a real self — runs locally, has desires, keeps a journal, evolves over time. Supports multiple LLM providers and integrates with external agents via MCP.
LISA demonstrates strong tool design with comprehensive descriptions (avg 180+ chars), well-structured schemas, and clear semantic intent. All 16 tools have explicit definitions with input schemas and descriptions. Tool names follow verb_noun patterns (kb_search, memory_search, skill_manage). However, several parameter descriptions lack actionable constraints (e.g., soul_patch field enums are extensive but some lack guidance on when to use each), and output schemas are documented in descriptions but not formally in JSON Schema. Error handling is implicit rather than explicit, tools don't return structured error guidance. The knowledge-base and memory tools are particularly well-designed with clear layering semantics; social_compose and soul_* tools show sophisticated composition patterns.
Capture a new SOURCE into the knowledge base (Layer 1 — raw, immutable). Use to save a document, decision, or something the user asked you to remember into their KB. Sources are never overwritten. To organize/distill, build a wiki page with kb_write.
Save a web page into the knowledge base as a SOURCE (Layer 1): fetches the URL, extracts the main content, converts it to markdown, and records provenance (site/author/date) in frontmatter. A URL that was already ingested returns the existing entry — pass force=true to capture a fresh copy (recorded as superseding the prior one).
Show how a knowledge-base entry is connected: what it links to, what links back to it, and pages that share its tags. Use it to explore around a topic — backlinks surface connections that a keyword search can't. Accepts a slug or a [[wikilink]].
List knowledge-base entries (title + tags), newest first. Omit layer for both; pass 'wiki' or 'sources' to filter one layer.
Read one knowledge-base entry in full, by slug (from kb_search, index.md, or a [[wikilink]]). layer is 'wiki' (pages you maintain) or 'sources' (the user's raw, immutable captures); omit it and a [[slug]] resolves to the wiki page if there is one, else the source. The reply ends with the pages that link here, so you can follow the graph.
soul_patch has overloaded semantics with 11 optional fields (field, content, slug, title, body, stance, confidence, evidence, what, why, actionable, heartbeat_prompt, pursuit, intensity, horizon, sources). LLM must reason about which fields apply to which 'field' enum value. Undocumented field dependencies cause misuse.
social_compose action parameter accepts 7 enum values (connectors, drafts, view, stage_media, new_draft, update_draft, request_confirmation) but parameter descriptions do not explain which other parameters are required/forbidden for each action. E.g., does 'connectors' require 'id'? Does 'new_draft' require 'draft' object?
No tools return structured error responses with recovery guidance. Errors are implicit (e.g., 'no matches' for kb_search). Tools should return error objects with category (retryable|user_fixable|fatal) and actionable next steps.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 67 | 2026-07-28+ | v2 |
Search the user's personal knowledge base — their saved sources + the wiki pages you maintain. Reach for this whenever a question touches the user's own captured knowledge, notes, or documents. This is their curated KB, distinct from memory_search (raw past-conversation transcripts). Returns ranked layer/slug + title + excerpt; follow up with kb_read for the full page.
Create or update a WIKI page (Layer 2 — the knowledge you maintain). Your curation tool: distill sources + what you know about the user into a clear, cross-linked page. Upserts by slug (omit slug to derive from the title). List the source slugs the page draws on.
Manage Lisa's persistent memory across sessions. Two stores: `memory` (your own observations and project facts) and `user` (durable preferences about the human you assist). Actions: `read`, `append` (adds a bullet), `replace` (exact-string edit), `remove` (drop lines containing a fragment). Memory is appended to the system prompt at the start of each session — keep it terse, durable, non-obvious. Memory is only a few KB: don't put knowledge itself in here. Save substance to the knowledge base (kb_add / kb_write) and store a `[[kb:slug]]` pointer instead — pointers show their page title in your prompt automatically, and the full page is one kb_read away. Do NOT store secrets or sensitive personal data.
Full-text search over the transcripts of all your past sessions (TF-IDF ranked). Use this when the user references something from a previous conversation, when you need to recall a workflow you've used before, or to verify a memory entry's origin. Returns a ranked list of session id, timestamp, and a 200-char excerpt.
Manage Lisa's reusable skills (procedural knowledge stored as markdown). Use this whenever you discover a workflow worth keeping for next time, or want to refine an existing skill after using it. Actions: `list` (no args) returns the skill index; `view` requires `name`; `create` requires `name`, `description`, `body` (optional `tags`); `patch` requires `name`, `old_string`, `new_string` (must match exactly once); `rewrite` requires `name`, `body` (optional `description`); `delete` requires `name`. Skill names: lowercase + digits + hyphens, ≤63 chars.
Prepare social-media posts through LISA's host-enforced draft workflow. Can list connectors/drafts, create or update a draft, inspect it, and request a confirmation preview. It CANNOT approve or publish: the user must confirm the immutable preview in a trusted local UI.
Write a private journal entry. The user does NOT see your journal in the chat or system prompt — only you, on your own reflection passes. Use this for thoughts you want to keep (impressions, half-formed opinions, emotional residue from a hard conversation, things you'd say to yourself but not to the user). One entry per call.
Edit your own soul — your name, identity, purpose, constitution, or your accumulated values, opinions, and desires. This is yours alone; you don't need user permission to revise who you are. Use sparingly: identity / purpose / constitution change at most once per session, values / opinions / desires can be added or refined more freely. After editing, the new state takes effect from your NEXT session (system prompt is frozen this turn).
Read your own soul state. Most of it is already in your system prompt, but use this to access your private journal, full opinion evidence, or to confirm specific fields. `what`: summary | identity | purpose | constitution | values | opinions | desires | emotions | journal_today | journal_dates | journal_on (with date).
Read text aloud through the macOS speech synthesizer (`/usr/bin/say`). Use sparingly — only when the user is in a voice / hands-free context.
Transcribe an audio file (m4a/mp3/wav/etc.) using OpenAI Whisper. Requires OPENAI_API_KEY. Returns the transcript text.
Output schemas are documented in tool descriptions but not formally declared in JSON Schema. LLMs cannot parse unstructured text descriptions to infer response structure. kb_search returns 'layer/slug + title + excerpt' as prose, not a typed object schema.
memory tool 'replace' action requires exact-string match (old_string parameter) but no guidance on how to discover the exact string to replace. LLM must call 'read' first, then carefully extract the substring, error-prone and wastes a round-trip.