Dimension 2: Deep Vendor Profiles¶
1. Honcho (honcho.dev)¶
- Docs: https://docs.honcho.dev
- GitHub: https://github.com/plastic-labs/honcho
- Architecture: Two-layer context injection — base layer (session summary + representation + peer card) plus dialectic supplement (LLM reasoning). Dialectic automatically selects cold-start prompts (general user facts) vs. warm prompts (session-scoped context).
- Three orthogonal config knobs: contextCadence, dialecticCadence, dialecticDepth (1-3 passes)
- Multi-peer setup: Workspace contains Peers (users + AI agents). Each AI peer builds independent representation/card from its own observations.
- Reasoning engine ("deriver"): Background reasoning processes messages to extract premises, draw conclusions, build representations.
- Tools (5): honcho_profile, honcho_search, honcho_context, honcho_reasoning, honcho_conclude
- Key insight: The only provider built specifically for user modeling — not just storing facts but building a behavioral model of the user over time.
- Cost: Paid cloud / free self-hosted
- License: AGPL-3.0
2. OpenViking (github.com/volcengine/OpenViking)¶
- Docs: No dedicated docs site; README-driven
- GitHub: https://github.com/volcengine/OpenViking
- By: ByteDance (Volcengine)
- Architecture: Hierarchical virtual filesystem with
viking://URI scheme. Tiered context loading: - L0 (Abstract): ~100 tokens, one-sentence summary for quick identification
- L1 (Overview): ~2K tokens, core info and usage scenarios for planning
- L2 (Detail): Full content, loaded only on demand
- Retrieval: Directory recursive retrieval — vector similarity to identify directory, secondary search within directory, drilling down recursively. Every step logged as visible trajectory.
- Auto-extraction: On session commit, extracts into 6 categories (profile, preferences, entities, events, cases, patterns)
- Tools (5): viking_search, viking_read, viking_browse, viking_remember, viking_add_resource
- Key insight: 80-90% token reduction — instead of loading all context, agent "skims" at L0, refines at L1, deep-dives at L2 only when needed.
- Cost: Free (AGPL-3.0)
- Deployment: Self-hosted only, requires Docker + LLM for extraction
3. Mem0 (mem0.ai)¶
- Docs: https://docs.mem0.ai
- GitHub: https://github.com/mem0ai/mem0
- Architecture: Hybrid triple-store — vector + key-value + knowledge graph. LLM pass extracts structured facts from conversations, stores across all three layers.
- Dual memory scope: Session memories (short-term) + User memories (long-term). Both searched and injected before each response.
- Server-side LLM extraction: Mem0's infrastructure decides what's worth keeping — no manual curation.
- Circuit breaker: If API fails 5 times, stops calling for 2 minutes. Agent keeps working.
- MCP Server: https://mcp.mem0.ai/mcp — cloud-hosted MCP server
- Tools (3): mem0_profile, mem0_search, mem0_conclude
- Memory types: Conversation memory, Session memory, User memory, Organizational memory
- Key insight: Best SDK quality, largest community (51.4K stars), self-host option, Apache license. Fastest 30-second setup.
- Benchmark: LongMemEval-S 67.6%
- Cost: Freemium (10K adds/1K recalls free), $19/mo Starter, $249/mo Pro (Graph)
- License: Apache 2.0
4. Hindsight (hindsight.vectorize.io)¶
- Docs: https://hindsight.vectorize.io
- GitHub: Part of Vectorize
- Architecture: TEMPR — four parallel retrieval strategies: Temporal, Entity, Metadata, and BM25 for exact keyword matches. Three-stage pipeline: retain (ingest) → recall (retrieve) → reflect (synthesize).
- Knowledge graph: Stores structured knowledge — discrete facts, named entities, relationships. NOT raw text chunks.
- Reflect operation: Unique cross-memory synthesis — periodic pass that reads across all stored memories to derive higher-level insights and consolidate related facts. No other provider has this.
- Local mode: Embedded PostgreSQL, no cloud account required. Full features locally for free.
- MCP Server: https://api.hindsight.vectorize.io/mcp/default/
- Tools (3): hindsight_retain, hindsight_recall, hindsight_reflect
- Key insight: Highest LongMemEval scores (91.4-94.6% with Gemini-3). Independently validated (Virginia Tech arXiv paper). Best for coding agents.
- Benchmarks: LongMemEval 91.4-94.6%, BEAM 64.1%, LoCoMo 89.6%
- Cost: Free (local) / Cloud: $15/M retain, $0.75/M recall, $3/M reflect
- License: MIT
5. Holographic (built into Hermes)¶
- Docs: Built into Hermes docs
- Architecture: Holographic Reduced Representations (HRR) algebra on local SQLite + FTS5. Zero external dependencies.
- Trust scoring: Asymmetric feedback — +0.05 helpful / -0.10 unhelpful. Facts have trust scores (0.0-1.0).
- Fact store tool: 9 actions — add, search, probe, related, reason, contradict, update, remove, list
- Unique capabilities:
probe— entity-specific algebraic recall (all facts about a person/thing)reason— compositional AND queries across multiple entitiescontradict— automated detection of conflicting facts- Tools (2): fact_store (9 actions), fact_feedback
- Key insight: Most private option — SQLite file on disk, zero network calls, zero dependencies. Operational in seconds with no accounts.
- Cost: Completely free
- License: MIT
6. RetainDB (retaindb.com)¶
- Docs: Limited public docs
- GitHub: https://github.com/RetainDB
- Architecture: Turn-by-turn extraction with 3-turn context. Atomic memories with eventDate, documentDate, confidence scores.
- Retrieval: Full chronological retrieval — provides complete memory chronology rather than lossy semantic search. The answering model sees the full timeline.
- Hybrid search: Vector + BM25 + reranking
- Delta compression: Memory diff storage
- Tools (5): retaindb_profile, retaindb_search, retaindb_context, retaindb_remember, retaindb_forget
- Key insight: 0% hallucination rate claimed, SOTA on preference recall (88%). "No lossy retrieval" philosophy — guarantees coverage but gives model more context to sift through.
- Cost: $20/mo (no free tier)
- License: Proprietary
7. ByteRover (byterover.dev)¶
- Docs: https://docs.byterover.dev
- Architecture: Hierarchical knowledge tree stored as human-readable Markdown files in
.brv/context-tree/. Navigable hierarchy. - Curation engine: Five operations — ADD, UPDATE, UPSERT, MERGE, DELETE. Agent decides how to modify the context tree.
- Pre-compression extraction: Fires specifically before Hermes compresses a long conversation, ensuring in-flight knowledge gets captured before context is summarized away.
- Tiered retrieval: Fuzzy text → LLM-driven search
- Connectors: MCP, Skill, Hook, Rules types for 20+ agents
- Tools (3): brv_query, brv_curate, brv_status
- Key insight: Full visibility into what the agent knows — human-readable, editable Markdown. Best LoCoMo score (92.2%). Local-first, portable.
- Benchmark: LoCoMo 92.2% (single-hop 95.4%, temporal 94.4%)
- Cost: Free (local) / $19/mo Pro (cloud sync)
8. Supermemory (supermemory.ai)¶
- Docs: https://supermemory.ai/docs
- GitHub: https://github.com/supermemoryai
- Architecture: Semantic understanding graph built on top of entities. Documents vs Memories distinction — documents are raw input, memories are intelligent knowledge units created by Supermemory.
- Memory relationships: Three types — Updates (information changes), Extends (information enriches), Derives (information infers)
- Context fencing: Strips recalled memories from captured turns to prevent recursive memory pollution
- Session-end conversation ingest: Richer graph-level knowledge building
- Multi-container mode: Read/write across named containers
- Connectors: Google Drive, Notion, OneDrive, Web Crawler, Gmail, GitHub, S3
- Tools (4): supermemory_store, supermemory_search, supermemory_forget, supermemory_profile
- Key insight: All-in-one philosophy — memory + RAG + user profiles + connectors. Most generous free tier (1M tokens + 10K searches).
- Benchmark: LongMemEval 81.6% (with GPT-4o)
- Cost: Free (1M tokens/10K searches), $19/mo Pro, $399/mo Scale