oleg.m/waggle-os
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
master
waggle-os/docs/research/04-competitive-landscape.md
Oleg Maslov 0c3e2ead3b
Some checks failed
Installer Smoke / installer-smoke (push) Has been cancelled
Details
moving
2026-09-02 10:10:29 +02:00

22 KiB
Raw Permalink Blame History

Competitive Landscape: hive-mind vs. the 2025-2026 AI Memory Market

Date: 2026-04-22 (initial); updated 2026-05-02 (Hermes Agent integration per strategy/competitive/2026-04-30-hermes-agent-intel-update.md) Scope: Research-only competitive positioning for hive-mind (local-first AI memory, MCP server, SQLite, FTS5+vector hybrid, knowledge graph, I/P/B frames, wiki compiler).

Naming disambiguation (added 2026-05-02):

  • "Hermes" in this document, when referenced as MCP client (Strengths §1.4, Opportunities, Tagline #3), refers to the Nous Research Hermes AI coding client — consumer of MCP servers, friendly to hive-mind (positive distribution channel).
  • "Hermes Agent" (added below as §1.11) refers to the Nous Research closed learning loop agent product launched 25 February 2026 — direct competitor in the OSS knowledge worker agent space.

1. Competitor Profiles

1.1 mem0 (mem0.ai)

  • Architecture: LLM-driven extract/consolidate/retrieve layer on top of pluggable stores (vector + optional graph + relational). Requires an LLM (default GPT-4.1-nano) and embedding model. Graph memory is a paid variant.
  • Benchmarks: 91.6 on LoCoMo, 93.4 on LongMemEval, 64.1/48.6 on BEAM 1M/10M, ~7k tokens per retrieval (token-efficient algorithm, 2025). [mem0.ai/research]
  • Local-first? Self-hostable (Apache-2.0), but fundamentally cloud-shaped: still calls OpenAI/embedding APIs by default; best features (graph memory) are managed-only.
  • MCP: Yes, first-party mem0 MCP server + OpenMemory MCP (local).
  • Pricing: Hobby free (10k memories), Starter $19, Pro $249 (graph + analytics), Enterprise custom. [mem0.ai/pricing]
  • License: Apache-2.0 (core) / proprietary (cloud).
  • Differentiator: Market leader on benchmark headlines; 41k+ GitHub stars; $24M Series A (Oct 2025).

1.2 Letta (formerly MemGPT)

  • Architecture: Block-based context memory. Three tiers: Core Memory (in-context RAM block the agent edits), Recall Memory (searchable conversation log), Archival Memory (tool-queried cold storage). Self-editing is the signature feature.
  • Benchmarks: No LOCOMO headline; legacy MemGPT paper established the "LLM as OS" metaphor.
  • Local-first? Self-hosted server runs locally; needs external LLM for agent reasoning. ADE connects to local or cloud.
  • MCP: Yes — Letta consumes MCP servers as tool sources; third-party Letta-MCP-server exists.
  • Pricing: Self-hosted free; cloud $20–200/mo. [letta.com]
  • License: Apache-2.0.
  • Differentiator: Stateful, self-editing agents — agent writes its own memory blocks.

1.3 Zep / Graphiti (getzep.com)

  • Architecture: Graphiti temporal knowledge graph (Apache-2.0) backing Zep Cloud. Bi-temporal edges (valid-from / valid-to). Typically requires Neo4j/FalkorDB/Kuzu. Zep arXiv 2501.13956 is the paper.
  • Benchmarks: Zep claimed 84% LoCoMo → mem0 re-evaluated as 58.44% → Zep counter-claimed 75.14% → Dec 2025 new claim of 80% at <200ms P95. Graphiti claims 94.8% on DMR and +18.5% on LongMemEval. Treat with salt — there is an open public benchmark dispute. [GitHub issue getzep/zep-papers#5]
  • Local-first? Graphiti OSS is self-hostable but heavyweight (Neo4j). Zep Cloud is cloud-only; Zep Community Edition was deprecated April 2025.
  • MCP: Yes — Graphiti MCP Server v1.0 shipped Nov 2025, 20k+ GitHub stars. [blog.getzep.com]
  • Pricing: Free 1k credits/mo (prototype only), Flex $25/mo, custom enterprise.
  • License: Apache-2.0 (Graphiti) / proprietary (Zep Cloud).
  • Differentiator: Temporal knowledge graph done seriously; best-in-class MCP adoption.

1.4 Cognee (cognee.ai)

  • Architecture: ECL pipeline (Extract → Cognify → Load). DataPoints = strongly-typed Pydantic objects acting as both node and edge schemas. Unifies relational + vector + graph. 38+ connectors. $7.5M seed.
  • Benchmarks: HotPotQA: 0.93 human-like correctness (Cognee 2025.1), beating LightRAG, Graphiti, mem0. Multi-hop strength.
  • Local-first? Yes — self-hostable with local LLMs via Ollama. MCP: Yes, first-party. Pricing: OSS free; Core/Enterprise = contact sales. License: Apache-2.0.
  • Differentiator: Strongest multi-hop graph reasoning; schema-typed knowledge.

1.5 Supermemory (supermemory.com)

  • Architecture: Cloudflare Durable Objects, per-user MCP via URL-path isolation, SSE real-time. "Human-like decay" model. MemoryBench OSS benchmark suite.
  • Benchmarks: 81.6% LongMemEval (GPT-4o); strong LoCoMo / ConvoMem. $3M Oct 2025.
  • Local-first? No — cloud-first SaaS. MCP server OSS, engine cloud-resident. MCP: Yes, one of the most widely integrated (Claude Desktop, Cursor, Windsurf, VS Code, Claude Code). Pricing: Dev: Free / Pro $19 / Scale $399. Consumer: Free / $9. License: OSS client / proprietary engine.
  • Differentiator: Best distribution across AI clients; consumer "second brain" app.

1.6 LlamaIndex Memory Modules

  • Architecture: Library abstractions — ChatMemoryBuffer (FIFO), ChatSummaryMemoryBuffer (periodic summarization), SimpleComposableMemory (deprecated), and the newer unified Memory class with short-term FIFO + optional long-term extraction.
  • Benchmarks: None — this is plumbing, not a memory system.
  • Local-first? Runs where you host it; no storage opinion. MCP: No first-party memory server. Pricing: Free OSS. License: MIT.
  • Differentiator: Embedded in the most popular RAG framework; low-ceremony for existing LlamaIndex users.

1.7 LangMem (LangChain)

  • Architecture: SDK over LangGraph's BaseStore + Checkpointers. Semantic / episodic / procedural memory types. Background manager extracts and consolidates asynchronously.
  • Benchmarks: No public LOCOMO numbers. Local-first? Library-level. MCP: Indirect only. Pricing: LangSmith tiers for managed observability. License: MIT.
  • Differentiator: Tight LangGraph integration; opinionated episodic/procedural/semantic split.

1.8 OpenAI ChatGPT Memory

  • Architecture: Proprietary; "saved memories" (facts) + "reference chat history" (retrieval over past conversations). Rolled to free tier June 2025.
  • Benchmarks: None published. Local-first? No. MCP: No — and critically not exposed via API. [memobase.io blog] Pricing: Bundled. License: Proprietary.
  • Differentiator: Default memory for ~800M consumer users; zero-setup.

1.9 Anthropic Claude Memory + Memory Tool

  • Architecture: Two products. (a) Consumer Claude Memory — GA Team/Enterprise Sept 2025, Pro/Max Oct 23 2025, free tier Mar 2 2026. Project-scoped, work-pattern-focused. (b) Memory Tool (API, beta header context-management-2025-06-27) — client-side file directory the model CRUDs. Combined with context editing: +39% vs baseline.
  • Local-first? Memory Tool is client-side by design — Anthropic gives the protocol, you host the files. Friendly to local storage. MCP: Anthropic authored MCP; Memory Tool is complementary. Pricing: Consumer bundled; Memory Tool = API tokens. License: Proprietary / open spec.
  • Differentiator: The memory protocol standard-setter.

1.10 Basic Memory (basicmachines-co)

  • Architecture: Persistent semantic graph stored as plain Markdown files, indexed in a local SQLite. Obsidian-compatible. MCP-native.
  • Benchmarks: None. Local-first? 100% yes — closest philosophical sibling to hive-mind. MCP: Yes, first-party. Pricing: Free OSS. License: AGPL-3.0 (more restrictive than hive-mind's Apache-2.0).
  • Differentiator: Obsidian interoperability; markdown-as-source-of-truth. Lowest lock-in competitor.

1.11 Hermes Agent (Nous Research) — added 2026-05-02

  • Architecture: Closed learning loop, prompt-augmented agentic workflows. Components: MEMORY.md + USER.md text files maintained per-project (prompt memory), SQLite FTS5 full-text-search storage of past sessions (episodic archive), auto-generated markdown skills capturing repeated workflows (procedural skills).
  • Launch date: 25 February 2026.
  • Benchmarks: Internal-only — claims 40% speedup on repeat tasks, not peer-reviewed, no public methodology. Has not engaged any standardized public benchmark venue (LoCoMo, LongMemEval, Gaia2, τ³-bench).
  • Local-first? Yes — files-on-disk + local SQLite. MCP: Self-published as agent harness (consumes MCP servers, plus emits its own memory protocol). Pricing: Free OSS. License: Apache 2.0 (permissive).
  • Adoption: ~110K GitHub stars 10 weeks post-launch — strongest OSS adoption velocity in this category in 2026 to date.
  • Differentiator: "Agent that gets better over time at your specific workflows" — functionally identical positioning to Waggle's self-evolution narrative. This is the only competitor in the local-first quadrant that explicitly markets the same architectural-philosophy story.
  • Threat level: MEDIUM-HIGH (architectural philosophy overlap is non-trivial risk; mitigation strategies in §3 Threats below).

2. Positioning Matrix

2.1 Axes: Local-first ↔ Cloud-first vs. Flat (chat/vector) ↔ Graph/Structured

Flat / Vector-only Graph / Structured
Pure Local-first Basic Memory; Claude Memory Tool; Hermes Agent hive-mind; Cognee (local mode); SuperLocalMemory
Hybrid / Optional LlamaIndex; LangMem; Letta (OSS) Cognee (hybrid); mem0 OSS (graph paid)
Pure Cloud-first Supermemory; OpenAI Memory; LangMem Cloud Zep Cloud; mem0 Cloud; ChatGPT Memory

Hermes Agent placement note (2026-05-02): Hermes Agent occupies the Pure Local-first + Flat/Vector cell (alongside Basic Memory and Claude Memory Tool) because its episodic archive is flat SQLite FTS5, not a graph. It does not contest the Pure Local-first + Graph/Structured quadrant where hive-mind sits, but it competes for the same buyer through narrative overlap rather than architectural overlap.

2.2 Axes: Benchmark-chasing ↔ Opinionated epistemology

Chases LOCOMO/LongMemEval Has a point of view on what memory is
Library / Framework LangMem, LlamaIndex Letta (block self-edit)
Graph-first system mem0, Zep/Graphiti Cognee (ECL+DataPoints); hive-mind (I/P/B)
Consumer/SaaS Supermemory ChatGPT Memory, Claude Memory
Markdown/file-native — Basic Memory

hive-mind sits almost alone in the local-first + graph + opinionated epistemology quadrant. Cognee is the only peer, but Cognee ships heavyweight pipelines and defaults to cloud LLMs; Basic Memory shares the local-first ethos but is flat markdown without bitemporal graphs or I/P/B framing.

3. SWOT — hive-mind

Strengths

  • True single-file local-first: SQLite workspace, sqlite-vec + FTS5 hybrid RRF, no cloud dependency, no Neo4j, no Cloudflare. Competitors either need an LLM server (mem0, Letta, Cognee) or a cloud (Zep, Supermemory, OpenAI, Anthropic consumer).
  • Opinionated epistemology (I/P/B frames): Intra=facts / Predicted=hypotheses / Bidirectional=corrections. No competitor has a comparable conceptual model — most collapse everything into "memories." This is a differentiating narrative.
  • Bitemporal knowledge graph: parity with Graphiti's headline feature, but without the Neo4j dependency.
  • MCP-native with 21 tools + first-class Claude Code / Codex / Hermes support — competitive with Supermemory (widest MCP distribution) and ahead of most (LangMem, LlamaIndex, OpenAI).
  • Harvest breadth: 11 adapters (ChatGPT, Claude, Gemini, Perplexity exports, PDF, MD, URL) is unmatched — mem0 and Zep have nothing equivalent; Basic Memory relies on manual markdown.
  • Wiki compiler: synthesis step nobody else offers as a first-class primitive.
  • Apache-2.0 — more permissive than Basic Memory's AGPL-3.0, matching mem0/Cognee/Letta/Graphiti.

Weaknesses

  • No LOCOMO/LongMemEval number published. In a market where the top-of-funnel conversation is benchmark X.Y%, this is a meaningful visibility gap. Even disputed numbers drive press (see Zep vs mem0 saga).
  • Smaller team / mindshare vs. funded players (mem0 $24M, Zep, Supermemory $3M, Cognee $7.5M, Letta well-funded).
  • No managed cloud option — some enterprises want a SaaS SKU. Currently 0% of that TAM.
  • Less mature SDK surface compared to LangChain/LlamaIndex ecosystem gravity.
  • I/P/B metaphor needs evangelism — video-codec framing is clever but requires education; mem0/Zep benefit from familiar vocabulary.
  • Extracted from a proprietary product (Waggle OS / Kvark) — perception risk around "is this really open or a freemium funnel?" needs managed narrative.

Opportunities

  • Local-first is the wedge. The 87% consumer privacy concern (Cisco 2025) + EU AI Act + enterprise data-residency requirements = a growing slice that refuses cloud memory. Basic Memory and SuperLocalMemory prove demand; hive-mind can own the category with a richer feature set.
  • Claude Memory Tool alignment. Anthropic's Memory Tool is client-side and storage-agnostic — hive-mind can position as the recommended local backing store for Claude Memory Tool. No competitor has staked that claim yet.
  • Benchmark publication. Running MemoryBench (Supermemory's OSS suite) or LoCoMo and publishing a credible number — even if 5-10 points under mem0's headline — converts hive-mind from "interesting" to "comparable."
  • Developer "AI coding memory" niche. Claude Code / Codex / Hermes MCP-first users skew technical and privacy-conscious. They don't want to send their codebase conversations to mem0 cloud. This is the highest-converting wedge.
  • Wiki compiler as a differentiator — nobody synthesizes frames into interlinked pages. This is a content product hiding inside a memory product.
  • Acquihire / strategic interest. Anthropic, GitHub, JetBrains, or Cursor could want a local memory layer. Being Apache-2.0 + single-file + MCP-native is acquirer-friendly.

Threats

  • Anthropic or OpenAI ships a default local memory bundled with their official client. ChatGPT/Claude desktop apps owning memory shrinks hive-mind's consumer TAM overnight.
  • mem0 or Supermemory ship a credible "local mode". Both have more eng resources; if mem0 bundles Ollama + sqlite-vec into a one-line install, the moat shrinks.
  • Graphiti / Zep pulls the graph-quality crown. If Graphiti ships a lightweight embedded backend (Kuzu is already embeddable), Zep could compete in the local-first quadrant.
  • LangChain/LlamaIndex gravitate to opinionated defaults — if either adopts mem0 or LangMem as the default and tutorials proliferate, greenfield devs never discover hive-mind.
  • Basic Memory or Cognee doubles down on local-first + graph and outpaces hive-mind in that quadrant.
  • Benchmark-war optics. The public Zep ↔ mem0 spat shows the category is noisy; entering without a number is risky, entering with a weak number is worse.
  • Hermes Agent (Nous Research) — added 2026-05-02. Architectural-philosophy overlap is non-trivial: Hermes Agent markets the same "agent that gets better over time" narrative, has 110K stars 10 weeks post-launch, and ships closed learning loop with prompt memory + SQLite FTS5 episodic archive + auto-generated procedural skills. However, six structural moats remain unaddressed by Hermes: (1) bitemporal knowledge graph vs flat SQLite FTS5, (2) I/P/B frame model with importance weighting vs undifferentiated text, (3) MPEG-4 frame compression + wiki compiler (neither in Hermes), (4) modular Apache 2.0 npm packages vs Hermes monolithic distribution, (5) EU AI Act audit triggers built-in vs not addressed, (6) peer-reviewed-style benchmark portfolio (LoCoMo apples-to-apples, GEPA cross-family, forthcoming Gaia2 + τ³-bench banking_knowledge) vs Hermes internal "40% speedup" only — this is the unbridgeable credibility moat for 2026. Probability Hermes engages public benchmark venue u 2026: ~30%. Probability Hermes builds bitemporal-style memory: low (architectural rewrite). Probability Hermes targets regulated industries: low (Apache hobbyist/dev market, no compliance positioning). Mitigation per strategy/competitive/2026-04-30-hermes-agent-intel-update.md §3: explicit differentiator messaging Day 0, head-to-head Gaia2 framing weeks 4-8, τ³-bench banking_knowledge framing weeks 8-12.

4. Positioning Taglines (5 options)

  1. "Your AI's memory. One file. Zero cloud." Minimalist, defensible, true. Owns the local-first quadrant in one line.

  2. "I, P, B — the only memory system that knows the difference between a fact, a guess, and a correction." Leads with the epistemology moat. Honest about what the I/P/B model buys you.

  3. "The memory layer for Claude Code, Codex, and Hermes — not for someone else's cloud." Targets the developer-MCP wedge directly; pits hive-mind against mem0-cloud and Supermemory-cloud.

  4. "SQLite in. Wiki out. Apache-2.0 all the way through." Honest-engineering vibe; contrasts with AGPL (Basic Memory) and proprietary cloud (Zep, OpenAI).

  5. "Graphiti-grade knowledge graph. Without the Neo4j. Without the bill." Directly attacks the graph leader's weakness (operational heaviness, cloud pricing).

5. Bottom Line

hive-mind's defensible position in the 2026 market is local-first + graph + opinionated epistemology + MCP-native + peer-reviewed-style benchmark portfolio — a quadrant currently contested only by Cognee (heavier, cloud-LLM-leaning) and Basic Memory (flat, AGPL, no graph). The market's benchmark arms race is noisy and partially discredited (Zep/mem0 dispute), but visibility still requires a number. Shipping a published LoCoMo/LongMemEval figure, aligning narratively with Anthropic's Memory Tool spec, and owning the "AI coding agent memory without the cloud" wedge are the three highest-leverage moves.

Updated 2026-05-02 — Hermes Agent (Nous Research) consideration: A new direct competitor in the local-first OSS knowledge worker agent space launched 25 February 2026 and has 110K GitHub stars 10 weeks in. Hermes Agent markets the same "self-improving agent" narrative as Waggle but ships flat SQLite FTS5 (vs hive-mind bitemporal graph), undifferentiated text storage (vs I/P/B framing), no benchmark engagement beyond internal "40% speedup" claim, and no regulatory positioning. Hermes Agent does not architecturally compete in hive-mind's quadrant but it does compete for the same buyer through narrative overlap. The defensible response is the peer-reviewed-style benchmark portfolio (LoCoMo apples-to-apples + GEPA cross-family + forthcoming Gaia2 + τ³-bench banking_knowledge) — that is the moat Hermes cannot match in 2026. All Day 0 launch messaging must explicitly include the six differentiators per strategy/competitive/2026-04-30-hermes-agent-intel-update.md §2.

Sources

Reference in New Issue View Git Blame Copy Permalink