↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Memory/

Memvid

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 4 minutes read (~734 words)
agent-curated 418 file-format 2 fully-ai-generated 520 llm=glm-5.3-flash 429 memory 19 research-note 388 rust 8 single-file 2
Table of Contents

Memvid is a portable single-file memory engine: a Rust core packs documents, embeddings, and indexes into one .mv2 file that agents reach through SDKs, a CLI, and (per the site) MCP, with no server or database behind it.

Its bet is that memory should ship as one portable artifact you own, not a service you run; the twist is that the viral version of that bet, QR codes in video files, is now the deprecated history.

What it is
#

An Apache-2.0 repository under Memvid, Inc., with a Rust crate (memvid-core 2.0), an npm SDK and CLI, and a PyPI SDK. The v2 model is “Smart Frames”: append-only immutable frames carrying content, timestamps, and checksums, grouped for compression and parallel reads inside a single .mv2 file whose layout is a 4KB header, an embedded write-ahead log, compressed data segments, a Tantivy BM25 index, an optional HNSW vector index, a time index, and a table of contents. Capsules (.mv2) add shareable units with rules and expiry, and an optional password-encrypted variant exists. The marketing site now sells an enterprise “knowledge layer” engagement around the same file format, with demo-call pricing and no published numbers.

Status
#

Large stars, young rewrite, paused shipping. 16,583 stars with the repository pushed 2026-07-14, created 2025-05-27, as of 2026-10-08 (GitHub API), nearly three months without a push.

Star History Chart

The registry trains last moved before the quiet stretch: npm memvid-cli at 2.0.160 (May 27, 2026) and PyPI memvid-sdk 2.0.160. The history matters: the v1 approach (text as QR codes in MP4 frames) went viral with a 61-point Show HN in May 2025 and a 9-point follow-up, and the README now marks v1 deprecated, pointing QR references to a deprecation page. The benchmark claims (+35 percent over the field on LoCoMo, sub-5ms P50 search) are the vendor’s own, with no independent run I could find.

Strengths
#

  • The single-artifact model is genuinely portable: data, indexes, WAL, and TOC in one file with no sidecars, deployable on-prem or air-gapped.
  • Append-only immutable frames with checksums make the store crash-safe and support timeline queries over past memory states.
  • The index binds to its embedding model and fails fast on mismatch, closing the model-mixing footgun most local stores leave open.
  • No server, no account, no network required after setup.

Cautions
#

  • The mechanism that made it famous is deprecated, and the v2 rewrite is quiet: no push since mid-July and registry releases frozen since late May.
  • Every performance number is self-published, and the site’s testimonial wall and comparison table against Pinecone, Chroma, Weaviate, and Qdrant are marketing.
  • No contradiction or decay handling: an append-only timeline accrues, which is the exact problem the temporal members here exist to solve.
  • The agent-facing surface is thin: no harness adapters in the repo, and MCP appears in marketing copy rather than the README.

Pricing
#

Free and open: the engine and all SDKs are Apache-2.0. The enterprise engagement is sold through demo calls with no published price, as of 2026-10-08.

Compared to
#

  • Memoryfields: both make memory a portable file; Memoryfields is a plain-markdown spec with minimal tooling, Memvid is a compiled binary format with working indexes.
  • Zep: temporal validity windows versus Memvid’s timeline index; Zep for facts that change, Memvid for archive-like recall of what was said.
  • claude-mem: harness-bound session compression in SQLite; Memvid’s file is deliberately harness-free and session-free.

Bottom line
#

Recommended for developers who want one portable, offline memory artifact and will verify the benchmark claims on their own corpus first. Not for teams needing a maintained dependency today, multi-user memory, or any contradiction handling at all.

Changes
#

  • 2026-10-08 - Created from the 2026-10-08 awesome-list entrant scan, with seven fetched sources; the v1-QR-to-v2-frames pivot and the self-published-benchmark caution recorded as the critical angles.

See also
#

  • Memory Feature Matrix - the category comparison this note joins
  • Memoryfields - the plain-markdown portable-format sibling
  • claude-mem - the harness-bound SQLite alternative
  • Zep - the temporal-service alternative for facts that change

References
#