RecallLong-term AI memory layer

A capsule is a memory asset you intentionally keep — a decision, thought, or key answer pulled out of an AI conversation worth saving long-term. It's distinct from a raw conversation:

Design philosophy: raw conversation is the process; capsule is the crystal. Six months later you don't want to find "what we discussed on March 14"; you want "why we picked Clerk over rolling our own auth" — that's the capsule.

Recall provides four entry points, covering manual / automatic / IDE flows:

The Recall page /dashboard/recall currently supports:

• Semantic search: Voyage-3 + pgvector cosine top-K — synonyms / cross-language / intent matching; auto-falls back to literal search during transient upstream outages (seamless to users)
• Literal search: ts_query full-text on title + content_md, 300ms debounce
• Tag filter: click any row tag chip to filter
• Auto Curator: auto-evaluates conversation segments worth keeping (rule scoring ~ms; gray-zone refinement via Haiku 4.5 on click)
• Chip recommendations: "Related memories" floating above chat input, semantically matched capsules
• Timeline view (Activity-band): density heatbar + top topic chips, toggle at top of Recall page
• Dedup hint: 7-day window cosine ≥ 0.85 → suggest merge before save
• Month grouping: auto-grouped by created_at YYYY-MM, collapsible
• Infinite scroll: 50 per page, IntersectionObserver auto-loads

Roadmap: Constellation (time × semantic 2D UMAP projection with auto-clustering + Haiku-generated cluster labels) — unlocks at ≥ 30 capsules. FIFO auto-overflow, weekly digest email, and cross-project project_id isolation arrive in later versions.

Recall is the only capacity-tiered subscription at Nexevo (Chat / Agent / Quorum themselves are token / task pay-as-you-go, no monthly fee). Three tiers Free / Pro / Max:

Capacity counts title + content_md + 8KB metadata overhead. Quota exceeded returns HTTP 429 quota_exceeded with upgrade link; per-capsule hard cap is 1MB (infrastructure protection), exceeded returns 413 capsule_too_large.

Admins can override per-user quota at /x_console/users/{user_id}(enterprise POC / support churn fix); set NULL to revert to tier default.

See /pricing.

Recall is exposed via the Nexevo MCP server to any MCP-aware client (Claude Desktop / Cursor / your agent). The server has one unified toolconductor_ask; Recall operations are selected via themode param:

• mode=save_memory — store a capsule
• mode=search_memory — semantic capsule recall

In Claude Desktop / Cursor, add the server config:

json
{
  "mcpServers": {
    "nexevo": {
      "command": "npx",
      "args": [
        "-y", "mcp-remote",
        "https://api.nexevo.ai/v1/mcp",
        "--header", "Authorization: Bearer YOUR_NEXEVO_API_KEY"
      ]
    }
  }
}

Once enabled, you can tell Claude:

"Save what we just discussed about Clerk; tag: auth, decision"

Claude auto-invokes conductor_ask (mode=save_memory) → hits /v1/memory/capsules REST endpoint → stored. Recall is symmetric: mode=search_memory.

Full MCP setup + all modes + Cursor config + FAQ: see/docs/mcp →

All endpoints use Bearer JWT auth with per-user isolation (cross-user access returns 404 to prevent existence leakage).

bash
# 创建胶囊 example
curl https://api.nexevo.ai/v1/memory/capsules \
  -H "Authorization: Bearer $NEXEVO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "title":       "Clerk 选型决策",
    "content_md":  "决定用 Clerk 而非自建...",
    "kind":        "decision",
    "tags":        ["auth", "decision"]
  }'