Engineer made built in memory, loop detection and audits mandatory for agents

Octopoda has launched an open-source memory and observability layer for AI agents that provides persistent memory, loop detection, audit trails, and a live dashboard. The tool installs via pip and works with frameworks like LangChain, CrewAI, AutoGen, and OpenAI Agents SDK, aiming to solve common agent failures such as memory loss, infinite loops, and lack of debugging visibility.

The open-source memory and observability layer for AI agents. Persistent memory, loop detection, audit trails, and a live dashboard — automatic on pip install . Website https://octopodas.com · · https://octopodas.com/docs Docs · https://octopodas.com/dashboard Dashboard · quick-start Quick start MCP server Live fleet overview: agent health, operations volume, per-agent scores, the anomaly stream, and the loops caught before they burned tokens. The same dashboard runs locally and in the cloud. What is Octopoda what-is-octopoda The problems it solves the-problems-it-solves Quick start quick-start Local vs cloud local-vs-cloud What you get out of the box what-you-get-out-of-the-box Agents agents · Memory memory · Shared memory shared-memory · Audit trail audit-trail Advanced features advanced-features Framework integrations framework-integrations MCP server mcp-server How it compares how-it-compares Cloud & pricing cloud--pricing Installation installation · Configuration configuration Octopoda is the layer between your AI agents and a production system that behaves. You write your agent however you like — plain Python, LangChain, CrewAI, AutoGen, the OpenAI Agents SDK, or MCP — and Octopoda sits underneath and handles four things agents consistently get wrong: Memory that survives every restart, crash, and deploy. Loop detection that flags a stuck agent in seconds, with the exact calls that caused it. An audit trail of every decision, write, and recovery — optionally hash-chained and verifiable. A live dashboard so you can actually see what your agents are doing. It runs locally with one pip install and zero infrastructure. When you outgrow local, the same code syncs to the cloud with a single environment variable — no re-architecture, no migration. The whole thing is MIT-licensed. If you have ever shipped an agent and watched it forget the user between sessions, loop on a failing API call, or vanish into a black box you couldn't debug, this is the missing layer. Agents forget on every restart. The moment your process restarts, the agent loses everything it knew about the user, the task, and the conversation. Octopoda gives every agent persistent memory that survives restarts, crashes, deployments, and kills — versioned by default. Agents loop, and quietly burn money. A stuck agent retrying a failing tool call can spend real money before anyone notices. Octopoda's detector catches retry, oscillation, ping-pong, reflection, and recall-write patterns in seconds and surfaces exactly which calls caused them. Detection is automatic on every write; intervention auto-pause, spend cap is opt-in through the v2 circuit-breaker config, so the policy stays yours. Agents are black boxes. When an agent does something surprising in production, you usually can't reconstruct why. Octopoda logs every decision, write, and recovery into a replayable timeline you can diff over time. Events written through the audit-v2 endpoint are hash-chained per agent prev hash → this hash , so you can verify integrity with a single call. Already have an agent on OpenAI, Anthropic, LangChain, CrewAI, AutoGen, or MCP? Add memory in two lines — no change to your agent's logic: pip install octopoda python import octopoda octopoda.init api key="sk-octopoda-..." the entire integration Octopoda auto-detects your framework, captures what matters from each turn, distills it into memories, and injects relevant recall into future calls — automatically. Or run any agent script unchanged from the terminal: export OCTOPODA API KEY=sk-octopoda-... octopoda-run python your agent.py auto-instruments on launch octopoda-run doctor checks your key + detected frameworks Get a free key at octopodas.com https://octopodas.com . Your agents and their memories appear on the live dashboard within about ten seconds of the first turn. Running multiple scripts that should share one brain? Set OCTOPODA AGENT ID=my-agent so they write to the same memory. On slow networks, raise OCTOPODA RECALL TIMEOUT=5 seconds . python from octopoda import AgentRuntime agent = AgentRuntime "my chatbot" agent.remember "user name", "Alice" kill the process. restart Python. then: print agent.recall "user name" .value 'Alice' — still there. Survives every restart, deploy, and crash. That is the whole setup. Your agent now has persistent memory, loop detection, crash recovery, and an audit trail. No config, no Docker, no Redis, no extra services. pip install octopoda server octopoda Open http://localhost:7842 — the same dashboard as the cloud version, running against your local data. No account, no API key. octopoda-init It walks you through pasting or signing up free for an API key, validates it, and saves it to ~/.octopoda/config.json . No environment variables to edit. The SDK auto-loads the key on the next import, and the same Python code above writes to the cloud and shows up live at octopodas.com/dashboard https://octopodas.com/dashboard . Prefer environment variables? export OCTOPODA API KEY=sk-octopoda-... Both methods work. The SDK checks the env var first, then the config file. Same Python API both ways. Start local; move to cloud when you need sync, team access, or the managed dashboard. | Local | Cloud | | |---|---|---| | Setup | pip install octopoda | Sign up free at octopodas.com | | Storage | SQLite on your machine | PostgreSQL + pgvector | | Dashboard | | octopoda ai extra ~33 MB OCTOPODA API KEY When you create an AgentRuntime , all of this runs in the background automatically — no configuration: | Feature | What it does | |---|---| | Persistent memory | Survives restarts, crashes, and deploys. Versioned by default. | | Loop detection | Five-signal engine: retry, oscillation, ping-pong, reflection, recall. | | Audit trail | Every write logged; audit-v2 events hashed and chained, replayable. | | Crash recovery | Automatic snapshots and heartbeat-based restore. | | Health scoring | Continuous per-agent performance and memory-quality monitoring. | | Goal tracking | Set goals and milestones per agent agent.set goal . | Every agent gets a live profile: score, operation count, read/write latency, spend, and loop-suppression stats. Drill into any agent for its latency trend, operation breakdown, timeline, memory, and checkpoints. Browse every memory an agent has written, filter by type fact, preference, summary, embedding , inspect version history, and see exactly how each value changed over time and which agent wrote it. agent.remember "user name", "Alice" agent.recall "user name" .value 'Alice' agent.recall history "user name" every prior version, newest first Memory is versioned automatically — each write appends a new version, and nothing is silently overwritten. Multiple agents working on the same problem can share knowledge through named memory spaces. Writes are atomic, reads are immediate, and every change is logged with its author — so you always know which agent contributed what. research agent.share "market size", "$2.1B AI memory market by 2027", space="team-knowledge" result = coding assistant.read shared "market size", space="team-knowledge" print result.value "$2.1B AI memory market by 2027" Spaces track authorship and timestamps for every write. Concurrent writes to the same key surface as a conflict last-write-wins by default via /safe ; use agent.shared conflicts space="team-knowledge" to review them. Every decision, crash, recovery, and anomaly is logged with full context — including a memory snapshot captured at the moment of the decision. Replay any time window and see exactly what each agent knew, decided, and why. agent.log decision decision="Keep single VPS instead of Kubernetes", reasoning="Current traffic doesn't justify K8s complexity.", context={"current rps": 14000, "threshold rps": 1000000}, Every log decision captures a memory snapshot at that instant, and a built-in similarity check warns you when a decision repeats a recent one. The timeline shows decisions alongside crashes and recoveries, filterable per agent. For tamper-evident provenance, write through the audit-v2 endpoints POST /v1/auditv2/event , GET /v1/auditv2/events . Those events are hashed and chained per agent prev hash → this hash ; GET /v1/auditv2/verify-chain returns ok=true plus a per-agent breakdown. The legacy log decision call writes a simpler row without the chain — route through audit-v2 when you need verifiable integrity. Everything below is optional. Reach for it when you need it. Semantic search — find memories by meaning, not exact keys agent.remember "bio", "Alice is a vegetarian living in London" results = agent.recall similar "what does the user eat?" Returns the right memory with a similarity score In cloud mode , embeddings are computed server-side and this works out of the box. In local mode , install the AI extra pip install octopoda ai so the local embedding model ~33 MB, CPU can run. Without it, recall similar returns 0 results locally and logs a warning. Agent messaging — agents talk through shared inboxes agent a.send message "agent b", "Found a bug in auth", message type="alert" messages = agent b.read messages unread only=True Goal tracking — goals and milestones per agent agent.set goal "Migrate to PostgreSQL", milestones= "Backup", "Schema", "Migrate", "Validate" agent.update progress milestone index=0, note="Backup done" Memory management — forget, consolidate, health agent.forget "outdated config" delete a specific memory agent.forget stale max age seconds=30 86400 clean up memories older than 30 days agent.consolidate dry run=False merge near-duplicates agent.memory health health report Snapshots & recovery agent.snapshot "before migration" ... something goes wrong ... agent.restore "before migration" Export / import bundle = agent.export memories new agent.import memories bundle Drop into the framework you already use. One line, and your agents get persistent memory. All integrations work locally no API key or with cloud sync OCTOPODA API KEY . LangChain — drop-in conversation memory python from octopoda import LangChainMemory memory = LangChainMemory "my-chain" memory.save context {"input": "I prefer dark mode"}, {"output": "Got it "} variables = memory.load memory variables {} CrewAI — persistent crew findings and task results python from octopoda import CrewAIMemory crew = CrewAIMemory "research-crew" crew.store finding "researcher", "market size", {"value": "$4.2B"} finding = crew.get finding "market size" AutoGen — multi-agent conversation memory python from octopoda import AutoGenMemory memory = AutoGenMemory "dev-team" memory.store message "user proxy", "assistant", "Research quantum computing" history = memory.get conversation history OpenAI Agents SDK — thread and run persistence python from octopoda import OpenAIAgentsMemory memory = OpenAIAgentsMemory memory.store thread state "thread 001", {"messages": ... } restored = memory.restore thread "thread 001" Give Claude, Cursor, or any MCP-compatible client persistent memory with zero code. pip install octopoda mcp Claude Code: claude mcp add octopoda -s user -e OCTOPODA API KEY=sk-octopoda-YOUR KEY -- python -m synrix runtime.api.mcp server Claude Desktop claude desktop config.json : { "mcpServers": { "octopoda": { "command": "python", "args": "-m", "synrix runtime.api.mcp server" , "env": { "OCTOPODA API KEY": "sk-octopoda-YOUR KEY" } } } } 28 tools for memory, search, loop detection, goals, messaging, decisions, and snapshots. A note on tool names the double prefix When you register the server as octopoda , the MCP client prefixes each tool with the server name. So the server-side octopoda remember is exposed to your agent as octopoda octopoda remember . That is correct client behaviour — just use the exposed name when you write skill files. If you register the server under a different name claude mcp add memory ... , the prefix changes to match. The full set of server-side tool names: octopoda remember · octopoda recall · octopoda search · octopoda recall similar · octopoda recall history · octopoda related · octopoda snapshot · octopoda restore · octopoda share · octopoda read shared · octopoda list agents · octopoda agent stats · octopoda process conversation · octopoda get context · octopoda log decision · octopoda forget · octopoda forget stale · octopoda memory health · octopoda consolidate · octopoda loop status · octopoda loop history · octopoda send message · octopoda read messages · octopoda broadcast · octopoda set goal · octopoda get goal · octopoda update progress · octopoda search filtered | Octopoda | Mem0 | Zep | LangChain Memory | | |---|---|---|---|---| | License | MIT | Apache 2.0 | Partial CE | MIT | | Local-first | Yes SQLite | Cloud-first | Cloud-first | In process | | Loop detection | Five-signal engine | — | — | — | | Agent messaging | Built in | — | — | — | | Audit trail | Hash-chained audit-v2 | — | — | — | | Crash recovery | Snapshots + restore | — | — | — | | Shared memory | Built in | — | — | — | | MCP server | 28 tools | — | — | — | | Semantic search | Local or cloud embeddings | Cloud embeddings | Cloud embeddings | Needs vector DB | | Framework integrations | LangChain, CrewAI, AutoGen, OpenAI Agents SDK | LangChain | LangChain | Own only | Sign up free at octopodas.com https://octopodas.com for the hosted dashboard, managed storage, and cloud API. python from octopoda import Octopoda client = Octopoda uses OCTOPODA API KEY agent = client.agent "my agent" agent.write "preference", "dark mode" results = agent.search "user preferences" | Free | Pro $19/mo | Business $49/mo | Scale $99/mo | | |---|---|---|---|---| | Agents | 5 | 25 | 75 | Unlimited | | Memories | 5,000 | 250,000 | 1,000,000 | 5,000,000 | | AI extractions | 100 | 10,000 | 50,000 | Unlimited | | Rate limit | 60 rpm | 300 rpm | 1,000 rpm | 5,000 rpm | | Loop detection | Basic | Full v2 | Full v2 | Full v2 | | Shared spaces | 1 | 5 | Unlimited | Unlimited | | Dashboard | Yes | Yes | Yes | Yes | | Support | Community | Email 48h | Priority | Dedicated | pip install octopoda Core — everything to get started Python 3.9+ pip install octopoda ai + local embeddings for semantic search pip install octopoda server + local dashboard server Flask pip install octopoda nlp + spaCy for knowledge-graph extraction pip install octopoda mcp + MCP server for Claude/Cursor Python 3.10+ pip install octopoda all everything Python 3.10+ Python versions.The core package supports Python 3.9+. The mcp extra needs 3.10+ the upstream mcp library does . On 3.9 and want everything except MCP? Use pip install octopoda ai,server,nlp . Local mode.Running without an API key gives you a fully working local install backed by SQLite at ~/.synrix/data/synrix.db . OCTOPODA API KEY accepts the sentinels local , offline , dev , none , or YOUR KEY HERE to force local mode explicitly. Real cloud keys start with sk-octopoda- ; anything else is treated as a local sentinel. Updating an MCP registration.If you change the claude mcp add octopoda ... env vars e.g. swapping local for cloud , restart the Claude Code window. A /mcp reconnect alone won't pick up new env, because the child process inherits Claude Code's cached env at startup. | Variable | Default | Description | |---|---|---| OCTOPODA API KEY | — | Cloud API key free at octopodas.com | OCTOPODA LICENSE KEY | — | License key for higher tiers optional | OCTOPODA LLM PROVIDER | none | openai , anthropic , or ollama | OCTOPODA OPENAI API KEY | — | Your OpenAI key for local fact extraction | OCTOPODA EMBEDDING MODEL | BAAI/bge-small-en-v1.5 | Local embedding model ~33 MB, CPU | SYNRIX DATA DIR | ~/.synrix/data | Local data directory SQLite + embeddings | OCTOPODA LOCAL MODE | unset | Set to 1 to force local mode regardless of the API key | SYNRIX HEARTBEAT INTERVAL SEC | 3 | Daemon heartbeat interval raise for low-resource boxes | SYNRIX MAX VERSIONS PER RUNTIME KEY | 10 | Cap on runtime: / metrics: key versions | The repo ships scripts under scripts/integration/ that exercise the product end to end against both a fresh PyPI install and live api.octopodas.com . Clone and rerun them: audit verify 3 1 13.py — live HTTP probes against production. mcp stdio harness.py — drives octopoda-mcp over JSON-RPC the way Claude Code does. user simulation.py — fresh venv, pip install octopoda from PyPI, exercises every SDK path. local dashboard smoke.py — proves the bundled dashboard serves byte-identical assets to the cloud one. See CONTRIBUTING.md /RyjoxTechnologies/Octopoda-OS/blob/main/CONTRIBUTING.md for setup and guidelines, and ROADMAP.md /RyjoxTechnologies/Octopoda-OS/blob/main/ROADMAP.md for what's planned. See SECURITY.md /RyjoxTechnologies/Octopoda-OS/blob/main/SECURITY.md for reporting vulnerabilities. MIT — use it however you want. See LICENSE /RyjoxTechnologies/Octopoda-OS/blob/main/LICENSE . Built by RYJOX Technologies https://octopodas.com · PyPI https://pypi.org/project/octopoda/ · Cloud API https://api.octopodas.com/docs · Dashboard https://octopodas.com/dashboard