Skip to content

Helix — Competitive Analysis

Status: Draft v1 · Last updated: 2026-06-18 · Related: PRD · Roadmap · Decisions

What this document is. A teardown of the AI-memory landscape as it stands mid-2026, scored against the six axes that define Helix's wedge. Star counts and pricing are point-in-time (mid-2026) and drift fast — treat them as directional, not authoritative. The thesis: every serious competitor optimizes for a hosted, server-side, latest-truth-wins memory service. Nobody ships a local-first, coding-native, single-file portable, git-like mergeable, $0-default memory layer. That gap is the entire reason Helix exists.


1. The market in one paragraph

There is no audited TAM for "AI memory" — every vendor estimate quietly bundles orchestration, RAG, and vector infrastructure, and the numbers don't reconcile (Mordor: $6.27B 2025 → $28.45B 2030; SkyQuest: → $69.13B 2033). The parent "AI agents" category is more defensible at ~$7.84B → $52.6B by 2030. But the qualitative trend is real and load-bearing: AI amnesia is the bottleneck. OpenAI shipped ChatGPT Memory GA in 2024 and expanded it April 2025; a16z's Big Ideas 2026 explicitly names context/state as the enterprise-agent bottleneck (https://www.a16z.news/p/big-ideas-2026-part-1). Meanwhile MCP has become the de-facto standard (Anthropic Nov 2024, OpenAI Mar 2025, Google Apr 2025) — which means MCP is table stakes, not differentiation. Speaking MCP gets you in the room; it does not win the room.


2. Per-product teardowns

Each teardown below scores memory model, retrieval, storage, portability, license/stars, pricing, the single biggest strength, and the gap Helix exploits.

2.1 Mem0 / OpenMemory

Memory model LLM 2-phase: extract → ADD/UPDATE/DELETE/NOOP, latest-truth-wins. Discrete facts, not a graph.
Retrieval Vector-first + BM25 / entity boosts.
Storage Qdrant default + SQLite history.
Portability OpenMemory MCP ran local via Docker but is being sunset for the self-host server; export is managed-only.
License / stars Apache-2.0 · ~58.8k ⭐ (the category leader by adoption).
Pricing Hobby $0 (10k) / Starter $19 / Growth $79 / Pro $249 / Ent. $24M total Series A, Basis Set, Oct 2025.
Biggest strength Ubiquity and ease — the default pip-install answer to "give my agent memory."
Gap Helix exploits Weak temporal reasoning (~49% LongMemEval vs Zep ~64%); lossy LLM extraction throws away source fidelity; no portable signed artifact, and export is gated behind the managed tier. Helix keeps the raw, owns the file, and signs it.

URLs: https://github.com/mem0ai/mem0 · https://mem0.ai/pricing

2.2 Letta (MemGPT)

Memory model OS-paging tiers — core (self-editable, in-context) / recall / archival. The agent edits its own memory via tools.
Retrieval Vector (text-embedding-3-small), no graph.
Storage Postgres + pgvector / SQLite / TurboPuffer.
Portability Agent File .af — open JSON of the whole agent, but unsigned, unencrypted, secrets not included.
License / stars Apache-2.0 · ~23.4k ⭐.
Pricing Self-host free; Cloud $20/mo + $0.10/agent + usage, BYOK. $10M seed, Felicis, Sept 2024, $70M post.
Biggest strength Durable, self-managing agents + the ADE debugger for inspecting agent state.
Gap Helix exploits Framework lock-in — Letta owns the agent loop, so memory comes bundled with their runtime. Python-only, costly context paging, no graph, no coding angle. The .af file is the right instinct executed weakly: unsigned plaintext JSON. Helix's signed + encrypted .dna leapfrogs it, and Helix is loop-agnostic (it's a memory layer, not an agent runtime).

URLs: https://github.com/letta-ai/letta · https://docs.letta.com/guides/build-with-letta/pricing

2.3 Zep / Graphiti

Memory model Bi-temporal knowledge graph (valid time + transaction time; facts are invalidated, not deleted). Episode / entity / community tiers.
Retrieval True hybrid: semantic + BM25 + graph BFS, with no LLM at query time (fast, deterministic, auditable).
Storage FalkorDB / Neo4j / Neptune.
Portability None — there is no portable artifact; you run a graph DB.
License / stars Graphiti Apache-2.0 ~27.6k ⭐ · Zep ~4.7k ⭐. The OSS Zep server is deprecated; self-host = the Graphiti library.
Pricing Graphiti free; Zep Cloud $0 → ~$104/mo cliff → ~$312 → Ent. ~$500K, YC W24.
Biggest strength Best-in-class temporal accuracy + audit trail + fast retrieval (the no-LLM-at-query-time design is genuinely excellent).
Gap Helix exploits You must run a graph DB — heavy, stateful, server-side. LLM-heavy ingestion. The turnkey self-host server is deprecated, pushing you to Cloud. No portable artifact. Helix steals the bi-temporal + no-LLM-at-query ideas without forcing a graph DB on a laptop.

URLs: https://github.com/getzep/graphiti · https://www.getzep.com/pricing/

2.4 Cognee

Memory model ECL pipeline (Extract-Cognify-Load), graph + vectors; memify self-prunes.
Retrieval Graph-vector hybrid, 14 search modes.
Storage Kuzu + LanceDB + SQLite default (many backends supported).
Portability Self-hostable, file-based + Cognee Cloud.
License / stars Apache-2.0 · ~17.9k ⭐.
Pricing $0 / Dev $35 / Team $200 / Ent. $7.5M seed, Pebblebed, Feb 2026.
Biggest strength Ontology-grounded graph + vector — the most "knowledge-engineering-correct" of the bunch.
Gap Helix exploits Multi-DB operational complexity (three+ stores to stand up), LLM-cost-heavy graph build, immature managed tier. Helix is a single signed file with zero infra to operate.

URLs: https://github.com/topoteretes/cognee · https://www.cognee.ai/pricing

2.5 LangMem

Memory model Semantic / episodic / procedural (self-updating prompts); hot-path + background extraction.
Retrieval Vector over a KV + embedding store.
Storage InMemory / Postgres / Pinecone / Redis.
License / stars MIT · ~1.5k ⭐.
Pricing OSS free; LangSmith Plus $39/seat.
Biggest strength Tri-type memory native to LangGraph.
Gap Helix exploits ~60s p95 latency (batch-only), LangChain coupling, low traction. Coupled to one orchestration framework; Helix is framework-neutral and synchronous.

URL: https://github.com/langchain-ai/langmem

2.6 Supermemory

Memory model Unified memory graph + a Memory Router drop-in (change one URL).
Retrieval Hybrid vector + keyword + graph, sub-300ms (vendor claim: "10x Zep, 25x Mem0").
Storage Cloudflare + Postgres (cloud) + an embedded ./.supermemory engine that runs offline with Ollama.
License / stars MIT · ~27.2k ⭐.
Pricing $0 ($5) / Pro $19 / Max $100 / Scale $399 / Ent. $2.6M seed, Oct 2025 (Susa, Browder, Jeff Dean angel).
Biggest strength Raw speed + drop-in router (lowest integration friction in the category) + a genuinely free local self-host.
Gap Helix exploits No standard export, self-host-at-scale needs the $399 tier, not coding-native, no signed artifact. The closest on "local + free" but it has no portable, verifiable memory object and no coding workflow.

URLs: https://github.com/supermemoryai/supermemory · https://supermemory.ai/pricing/

2.7 Memobase

Memory model Structured per-user profile (schema attributes) + an event timeline; buffered batch extraction.
Retrieval SQL profiles (<100ms) + vector events. Not a graph.
Storage Postgres + Redis.
License / stars Apache-2.0 · ~2.8k ⭐.
Pricing Self-host free + cloud PAYG.
Biggest strength Strong temporal numbers — 75.8% LoCoMo vs Mem0 66.9%, with cheap SQL-fast profile reads.
Gap Helix exploits Rigid schema, weak multi-hop reasoning, no graph, no export. Great if your memory fits a fixed profile shape; useless for open-ended coding context.

URL: https://github.com/memodb-io/memobase

2.8 A-MEM (feature source, not a competitor)

Memory model Zettelkasten notes + autonomous link generation + memory evolution (new notes update old ones).
Retrieval Vector + agentic linking.
Storage ChromaDB.
License / stars MIT · ~1.1k ⭐ (arXiv 2502.12110, Feb 2025).
Biggest strength The most adaptive / self-organizing memory in the field.
Gap Helix exploits Research-grade — no infra, no sync, no auth. We treat A-MEM as a feature source, not a rival: its self-evolving link generation is a capability to steal (see §4).

URL: https://arxiv.org/abs/2502.12110

2.9 basic-memory (closest philosophical rival)

Memory model Markdown-as-database — frontmatter entities + observations + relations; Obsidian wikilinks form the graph. Humans and AI edit the same files.
Retrieval SQLite FTS + FastEmbed vectors + graph traversal.
Storage Plain Markdown → SQLite index; cloud Postgres + Tigris S3.
Portability Best portability in the category — you own the Markdown; sync via Git or Syncthing; Obsidian-native; broad MCP support.
License / stars AGPL-3.0 · ~3.3k ⭐.
Pricing Self-host free; cloud $15/mo.
Biggest strength True file-ownership and human/AI shared editing — philosophically the closest to Helix.
Gap Helix exploits File-indexing scales poorly; AGPL-3.0 deters commercial embedding (an agent vendor cannot ship it inside a closed product); no encryption / signing; not coding-specialized; everything is plaintext. Helix wins on a signed + encrypted single artifact, coding-native modeling, git-like semantic merge (not just file-level Git sync), and a permissive license an agent vendor can actually embed.

URL: https://github.com/basicmachines-co/basic-memory

2.10 txtai (architecture model for .dna)

Memory model All-in-one embeddings database (vector + graph + SQL + keyword).
Portability Single compressed portable archive via save() / load() — the literal model for the .dna artifact.
License / stars Apache-2.0 · ~12.7k ⭐.
Gap Helix exploits SQLite / Faiss core not built for distributed scale; not framed as agent memory. We borrow the single-archive portability pattern and wrap it with signing, encryption, and coding-native semantics.

URL: https://github.com/neuml/txtai

2.11 Walrus + MemWal (Mysten / Sui)

Memory model Decentralized blob store (RedStuff erasure coding, ~4–5x replication; blobs are Sui objects). Base layer = blob read only, no semantic search.
Retrieval MemWal SDK (May 2026) adds encrypted memory containers + semantic search + Sui access control on top.
License / stars Apache-2.0 · ~375 ⭐ · WAL token.
Biggest strength Cheap, verifiable, content-addressed, vendor-neutral, on-chain ownership (~450TB).
Gap Helix exploits Storage-only core; MemWal is beta; the blockchain conflicts directly with $0 + local-first (a token requirement is a non-starter for a laptop-default dev tool). Takeaway: adopt content-addressing + encryption + verifiability, drop the chain.

URL: https://github.com/MystenLabs/walrus


3. The 6-axis positioning table

Six axes define the wedge. The claim is simple and falsifiable: no competitor checks all six.

Product Local-first Portable single-file Coding-native Git-like (semantic) merge MCP $0 default
Helix
Mem0 / OpenMemory ⚠️ (sunsetting) ❌ (export gated) ⚠️ (10k cap)
Letta (MemGPT) ⚠️ ⚠️ (.af unsigned) ✅ (self-host)
Zep / Graphiti ❌ (graph DB) ⚠️ (server deprecated)
Cognee ⚠️ (multi-DB)
LangMem ⚠️ ⚠️
Supermemory ⚠️ ($399 at scale)
Memobase ⚠️
A-MEM ⚠️
basic-memory ⚠️ (plaintext MD) ⚠️ (file-level Git)
txtai ✅ (archive)
Walrus / MemWal ❌ (chain) ⚠️ (blob) ❌ (token)

Legend: ✅ full · ⚠️ partial / conditional · ❌ none.

Read-out. basic-memory and Supermemory get closest on local-first + $0, but neither is coding-native, neither does semantic merge, and neither ships a signed/encrypted single artifact (basic-memory is plaintext under AGPL; Supermemory has no export). Letta's .af is the only true portable artifact among the leaders, and it's unsigned plaintext JSON bundled to a Python agent loop. The all-six column is empty except for Helix.


4. Helix's whitespace

The defensible whitespace is the intersection of three things no competitor combines:

  1. Coding-native memory. Memory modeled around code work — files, symbols, decisions, repo context, agent task history — not generic "user facts." Every competitor models a chat user; none models a coding agent's working set.
  2. Git-like semantic merge. Not file-level Git sync (basic-memory) and not latest-truth-wins overwrite (Mem0). A three-way semantic merge of memory: diff, conflict-detect, and reconcile memory the way you reconcile code. This is what makes the "review team memory like a PR" workflow possible (diff → approve → revert).
  3. Signed + encrypted single-file portability. The .dna artifact: one file you own, that is content-addressed (à la Walrus, minus the chain), encrypted at rest, and cryptographically signed so its provenance is verifiable. Letta's .af is the closest prior art and it is unsigned, unencrypted, and loop-bound.

No competitor occupies this intersection. The leaders are racing toward hosted server-side memory-as-a-service; the whitespace is the local, coding, mergeable, portable corner they're all walking away from.


5. Ideas to steal

Good artists copy; this section is explicit about what to lift and from whom.

Idea From What to take What to drop
Self-evolving links A-MEM Autonomous link generation + memory evolution (new notes retroactively update old). The research-grade ChromaDB-only plumbing; no infra/sync/auth.
No-LLM-at-query-time + bi-temporal Zep / Graphiti Deterministic, fast, auditable retrieval; facts invalidated (valid + transaction time) not deleted. The mandatory graph DB and LLM-heavy ingestion.
Content-addressing + encryption + verifiability Walrus / MemWal Content-addressed, encrypted, signed, verifiable memory blobs. The blockchain — a token requirement breaks $0 + local-first.
Single portable archive txtai save() / load() one-file archive pattern as the model for .dna. The "embeddings DB, not agent memory" framing.
File ownership + human/AI co-edit basic-memory You own the artifact; humans and agents edit the same memory. AGPL (un-embeddable), plaintext (no signing/encryption), file-indexing scale ceiling.
Drop-in integration ergonomics Supermemory One-line "Add to Cursor / Claude Code" install; near-zero integration friction. Cloud-router dependency and the $399 self-host-at-scale gate.

6. Bottom line

MCP is table stakes — every product here either has it or will. The category is consolidating around hosted, server-side, latest-truth-wins memory. Helix wins by refusing that frame: a local-first, coding-native, semantically mergeable, signed + encrypted single-file ($0-default) memory layer that an agent vendor can embed under a permissive license. The six-axis table has exactly one full row. That is the entire bet.


Sources

  • Mem0 / OpenMemory — https://github.com/mem0ai/mem0 · https://mem0.ai/pricing
  • Letta (MemGPT) — https://github.com/letta-ai/letta · https://docs.letta.com/guides/build-with-letta/pricing
  • Zep / Graphiti — https://github.com/getzep/graphiti · https://www.getzep.com/pricing/
  • Cognee — https://github.com/topoteretes/cognee · https://www.cognee.ai/pricing
  • LangMem — https://github.com/langchain-ai/langmem
  • Supermemory — https://github.com/supermemoryai/supermemory · https://supermemory.ai/pricing/
  • Memobase — https://github.com/memodb-io/memobase
  • A-MEM — https://arxiv.org/abs/2502.12110
  • basic-memory — https://github.com/basicmachines-co/basic-memory
  • txtai — https://github.com/neuml/txtai
  • Walrus / MemWal — https://github.com/MystenLabs/walrus
  • MCP — https://www.anthropic.com/news/model-context-protocol
  • a16z Big Ideas 2026 — https://www.a16z.news/p/big-ideas-2026-part-1

Star counts and pricing are mid-2026 point-in-time snapshots and change frequently; re-verify before quoting externally.