Skip to main content

20,000 Stars: How Hindsight Got Here, Version by Version

· 7 min read
Ben Bartholomew
Ben Bartholomew
Hindsight Team

Hindsight at 20,000 stars: a release timeline from the first open-source commit in December 2025 to v0.9.1

We open-sourced Hindsight in December 2025. Nine months and 67 releases later, it just crossed 20,000 GitHub stars. Across those releases we shipped more than 900 tracked changes — new capabilities, integrations, and hardening. The stars are the side effect. The releases are the story, so here is the real one: how an agent-memory engine went from a first public commit to what it is today, version by version.

The timeline at a glance​

VersionWhenReleasesThe headline
0.1.xDec 202513Foundations: embedded Postgres, local MCP, hindsight-embed
0.2.xJan 20261Multi-bank memory and cross-bank MCP tools
0.4.xFeb – Mar 202623The learning layer: observations, mental models, the knowledge graph
0.5.xApr 20267Faster retrieval, the Constellation view, template hub
0.6 – 0.7.xMay 20266Scaling the search layer, enterprise backends
0.8.xJun – Aug 20267Production hardening + a wave of coding-tool integrations
0.9.xAug 20262Knowledge Pages and memory for every coding agent

December 2025 — v0.1.x: the foundation​

Thirteen releases in the first month set the core that has not changed since: retain, recall, and reflect on embedded PostgreSQL, no external vector database to run. We moved fast on the local ergonomics so anyone could try it in minutes.

Highlights:

  • A local MCP server so any MCP client connects without a separate service, and hindsight-embed to run memory in-process.
  • An extensions system for plugging in new operations, plus the first graph-based retriever to use relationships between memories.
  • Memory banks, memory tags, backup/restore, and multilingual content support.
  • Early breadth on models and providers: LiteLLM, Cohere embeddings/reranking, configurable embedding dimensions, and Gemini 3 Pro / GPT-5.2.

January 2026 — v0.2.x: multi-bank memory​

A focused release that made isolation real. Multi-bank access plus the MCP tools to work across banks meant per-user and per-project memory stopped being theoretical. It also added Anthropic Claude and LM Studio, custom entities on retain, structured reflect output, and the first graph visualization in the Control Plane.

February to March 2026 — v0.4.x: the learning layer​

This is the big one, and the release line the earlier draft of this post undersold. Twenty-three releases and roughly 300 changes. If 0.1 gave Hindsight a place to put memories, 0.4 gave it the ability to learn from them.

The learning layer:

  • Observations — consolidated, deduplicated beliefs derived from raw facts — with configurable scopes, per-scope limits, history tracking, and batch consolidation.
  • Mental models — standing answers that refresh themselves — with full MCP create/read/update/delete, tag-aware triggers, staleness signals, and a history diff view.
  • A real knowledge graph: entity resolution, richer entity labels, and improved causal link detection, alongside temporal ordering.

Everything else scaled up around it:

  • Ingest anything: PDFs, images, and Office documents, the Iris parser, and async batch retain via provider Batch APIs.
  • Retrieval at scale: per-bank HNSW indexes, pgvectorscale (DiskANN) support, and a 40x speedup in observation recall on large banks.
  • Multi-tenancy and auth: Bearer-token MCP, hierarchical config scopes, a Supabase tenant extension, and consolidation.completed / retain.completed webhooks.
  • The integration wave began: Claude Code, Codex CLI, LangGraph, the Vercel AI SDK, Chat SDK, CrewAI, PydanticAI, AG2, Agno, Strands, and LlamaIndex, plus Go and AI SDK clients and native Windows support.

April 2026 — v0.5.x: retrieval grows up​

Seven releases, ~137 changes, focused on making recall faster and memory portable. The graph story consolidated onto a single LinkExpansion retriever, and a 3-phase retain pipeline dramatically improved ingestion throughput under concurrency.

Highlights:

  • The Constellation view — an interactive, zoomable canvas of the entity graph with heat-gradient coloring.
  • Delta mental-model refresh that only re-reads memory created since the last run.
  • Bank template import/export via a Template Hub: export a bank's config, mental models, and directives as a reusable manifest.
  • Local and open inference: a built-in llama.cpp provider, plus OpenRouter and first-class DeepSeek support.
  • New integrations: OpenAI Agents SDK, AutoGen, Paperclip, OpenCode, and Pipecat voice.

May 2026 — v0.6.x and v0.7.x: scaling the search layer​

Two minor lines, one theme: make retrieval fast and correct at real data sizes, and reach enterprise backends.

Highlights:

  • BM25 search backends: ParadeDB pg_search (Citus-compatible) and PGroonga for better multilingual search, plus configurable tokenization.
  • Enterprise storage: AlloyDB ScaNN indexing, an Oracle Database backend, and a read-replica option for recall traffic.
  • Smarter scheduling: bank-priority consolidation and targeted consolidation by observation scope.
  • More providers and models: z.ai, Fireworks, a litellmrouter with automatic fallback chains, and the Qwen3 reranker.
  • Integrations: Dify, n8n, SmolAgents, AWS Bedrock AgentCore, Google ADK, Flowise, Roo Code, Vapi voice, and Gemini Spark.

June to August 2026 — v0.8.x: built to run in production​

The second-biggest line in the project's history: seven releases and over 230 changes. This is where Hindsight got operationally serious, and where the integration list exploded.

Production hardening:

  • Periodic background maintenance that reconciles consolidation state and enforces retention across tenants on its own.
  • Durable, resumable progress snapshots for long-running consolidation and batch retains.
  • Reversible memory curation — edit, invalidate, and revert memory units — and semantic deduplication of near-duplicate observations.
  • A self-diagnosing health probe that reports whether the event loop is blocked or the connection pool is exhausted.
  • Whole-bank and cross-bank export/import (no re-running the LLM), plus prompt-prefix caching and Anthropic batch + prompt caching to cut cost.
  • Security: Memory Defense SIEM enrichment and multi-LLM failover / round-robin.

The integration wave:

  • GitHub Copilot (CLI and VS Code), Aider, Zed, OpenHands, Continue.dev, Cursor (plugin and CLI), Cline, Windsurf, Composio, Zapier, Obsidian, Haystack, the Microsoft Agent Framework, and more.

August 2026 — v0.9.x: Knowledge Pages, and memory for every coding agent​

The two most recent releases turn all of that infrastructure outward.

  • Knowledge Pages turn a bank into a self-healing wiki it writes about itself: living documents synthesized from consolidated memory that refresh as the bank learns.
  • Hindsight Coding Agents brings long-term memory to ten coding agents — Claude Code, Codex, Cursor CLI, opencode, Copilot CLI, and more — from one package, with per-bank toggles to tune temporal search, graph expansion, and reranking during recall.
  • 0.9.1 followed a week later: roughly 9x faster temporal extraction with the same results, whole-bank transfers that carry the Knowledge Pages tree, xAI OAuth for running the LLM on a SuperGrok subscription, and a database-independent liveness probe.

The throughline​

Read the timeline back and the pattern is clear. Nothing here is a prompt trick. It is databases, retrieval strategies, a learning layer, consolidation, ingestion, auth, and operational plumbing: the unglamorous infrastructure that makes memory accurate enough to measure and boring enough to run in production. More than nine hundred changes across sixty-seven releases, and the shape of the thing keeps getting sharper.

We should stay honest: Hindsight is not the most-starred project in the category, and the benchmarks are where we would rather compete anyway. If you have not tried it, start free on Hindsight Cloud or self-host in one command from GitHub. Thank you to everyone who shipped a release, filed an issue, or built something on top. The next version is already in progress.

Learn more: