celiums/celiums-memory

What's new

The biggest OSS expansion since v1.0. The engine grows from 6 tools to 20, gains a generic bring-your-own-key LLM client, and ships the full Ethics Engine (layers A + B + C). Zero proprietary endpoints, zero hardcoded model names, zero hidden fees — set up an OpenAI-compatible endpoint of your choice and everything works.

Highlights

🆕 20 MCP tools (was 6)

| Group | Tools | Requires |
|---|---|---|
| Knowledge | forage, absorb, sense, map_network | nothing — works offline with the bundled corpus |
| Memory | remember, recall, synthesize, bloom, cultivate | nothing — local store |
| Journal (new) | journal_write, journal_recall, journal_arc, journal_introspect, journal_dialogue | LLM (BYOK) |
| Write (new) | write_project_create/get, write_character_create, write_scene_create/update, write_continuity_check, write_export | LLM (BYOK) for continuity_check |
| Research (new) | research_project_create/list/continue, research_finding_add, research_gap_add, research_search, research_synthesize, research_export | LLM (BYOK) + optional corpus-search backend (CELIUMS_SEARCH_URL) |

🆕 BYOK LLM client (packages/core/src/llm-client.ts)

A 140-line OpenAI-compatible HTTP client. Set 1-3 env vars and the engine talks to whichever provider you prefer:

# OpenAI (default)
export CELIUMS_LLM_API_KEY=sk-...

# Ollama (local, free)
export CELIUMS_LLM_BASE_URL=http://localhost:11434/v1
export CELIUMS_LLM_API_KEY=ollama
export CELIUMS_LLM_MODEL=llama3.2

# OpenRouter / Together / Groq / vLLM / LM Studio — same pattern

Tools auto-gate themselves on CELIUMS_LLM_API_KEY. If you don't set it, AI-backed tools simply aren't registered — tools/list returns only OpenCore. The engine boots clean with nothing but a database.

🆕 Ethics Engine layers B + C

Layer A (semantic harm detection) was already public. v1.2 ships:

• Layer B — CVaR-probabilistic risk scoring (Conditional Value at Risk for content harm distributions)
• Layer C — 5-framework philosophical evaluation (utility / virtue / deontology / care / rights)

The full ethics gate now runs end-to-end before any memory is persisted.

🆕 Integration utilities

• integrations/crypto.ts — AES-256-GCM helpers for encrypting access tokens at rest
• integrations/schema.ts — tenant_integrations SQL schema (Notion/Slack/GitHub/Linear-ready)
• integrations/humanize.ts — opportunistic LLM-powered output formatting (markdown polish for inline rendering)
• integrations/intent-classifier.ts — turn free-form user queries into concrete MCP tool calls

🧹 OSS cleanup

• All proprietary endpoints removed from public code (atlas.celiums.ai, router.celiums.ai, etc.)
• All hardcoded model names removed from public docstrings
• .gitignore updated to keep proprietary tools out of the public repo

Migration

No breaking changes for existing users. v1.1 behavior is preserved end-to-end — just more tools available if you opt in to BYOK.

If you want the new AI-backed tools, set CELIUMS_LLM_API_KEY (and optionally CELIUMS_LLM_BASE_URL / CELIUMS_LLM_MODEL). See README → Configure your LLM (BYOK).

Full changelog

• feat(integrations): add encryption + schema utilities (36abac7)
• feat(mcp): add journal + write tools + generic LLM client (76461ca)
• feat(integrations): humanize + intent-classifier (BYOK, no proprietary deps) (17b4422)
• feat(mcp): add research tools (BYOK corpus + LLM) (17b4422)
• docs: README v1.2 — BYOK setup + 20 tools + Configure your LLM section (cead011)
• chore(mcp): drop hardcoded model names from docstrings (b24c637)

🍄 The engine just got a much richer surface, and stayed Apache 2.0 all the way through.