---
title: "Hindsight Memory for Hermes Agent: Is It Unlimited and Free?"
date: "2026-10-05"
video_id: "7693310660343516429"
tiktok_url: "https://www.tiktok.com/@josephbchandler/video/7693310660343516429"
creator: "@6530307856653112320"
category: "AI"
tags: ["hindsight", "hermes agent", "agent memory", "mcp", "vectorize"]
excerpt: "Hindsight adds long-term memory to Hermes Agent and other MCP clients. What the no-limit and free claims leave out, what Cloud costs and how it fits with Hermes."
---
Hindsight is an MIT-licensed long-term memory system for AI agents from Vectorize, and it plugs into Hermes Agent as an external memory provider. A TikTok from @josephbchandler says it is free, has "no limit," and fixes Hermes condensing its built-in memory. Self-hosting is free and raw storage is unbounded, but Hindsight condenses memories on purpose, its Cloud version is metered, and it runs alongside Hermes' capped memory files instead of replacing them.

## What is Hindsight?

The [Hindsight overview](https://hindsight.vectorize.io/) describes three operations (retain, recall and reflect) and a tiered memory structure: world facts, experience facts, observations, mental models and knowledge pages. Observations are consolidated beliefs drawn from raw facts, and mental models are curated summaries drawn from observations.

The [repository](https://github.com/vectorize-io/hindsight) is MIT licensed and ships a Docker image, a pip install, a Helm chart and a Python embedded mode.

## Does Hindsight give Hermes unlimited memory?

The video says, "Hermes does come with its own built-in memory, but as it fills up, the agent has to condense or remove things. Hindsight is better. There's no limit."

Hindsight's own blog disagrees. [The Consolidation Problem in Agent Memory](https://hindsight.vectorize.io/blog/2026/05/21/agent-memory-consolidation) opens with "An agent that remembers everything is an agent that remembers nothing useful." It lists four levers that every consolidation policy uses: importance, merge, decay and eviction. Hindsight runs all four.

One part of the claim holds. Hindsight says "eviction is a compliance tool, not a performance tool," used for GDPR, user deletion requests and PII redaction. It does not delete raw facts to free space. Raw storage is effectively unbounded, while the material the agent reads back is compressed and decayed by design.

## How does Hermes built-in memory work?

The settings screen in the video shows a "Memory Budget" of 2200 and a "Profile Budget" of 1375. The [Hermes memory docs](https://hermes-agent.nousresearch.com/docs/user-guide/features/memory) define these as character limits: `memory_char_limit: 2200` (roughly 800 tokens) and `user_char_limit: 1375` (roughly 500 tokens). Both are the stock defaults.

The two files live under `~/.hermes/memories/` and are loaded into the system prompt at session start. They are kept small so the prompt prefix stays stable for caching.

Hermes does not trim memory in the background. When the files are full, the write tool "returns an error instead of silently dropping entries," and the agent "makes room itself," consolidating or removing entries in the same turn before retrying.

## Does Hindsight replace Hermes memory?

No. The [Hermes memory providers docs](https://hermes-agent.nousresearch.com/docs/user-guide/features/memory-providers) say external providers run alongside built-in memory and only one external provider can be active at a time.

Hindsight's [Hermes integration guide](https://hindsight.vectorize.io/sdks/integrations/hermes) warns that with both active "the model may prefer the built-in one." It recommends disabling MEMORY.md and USER.md if you want Hindsight as the main store. If you keep them, "every entry the agent adds or replaces in them is also saved to the bank, tagged `builtin-memory`."

Installing Hindsight keeps a copy of what Hermes would otherwise prune, but the capped files keep getting pruned. In the video, the Hermes "Memory Provider" setting still reads "Built-in."

## Is Hindsight free?

The self-hosted build is. The speaker runs his instance on a Tailscale address, which means he self-hosts. The MIT license costs nothing, but every retain and reflect calls a language model, so inference still costs money unless you run a local model. The project documents `HINDSIGHT_API_LLM_PROVIDER=ollama` for that. The video does not say which model provider the speaker uses.

Hindsight Cloud, shown on screen behind a "Start free on Cloud" button and a $5 signup credit, is metered. The [billing page](https://docs.hindsight.vectorize.io/billing/) lists:

- Retain: $10.00 per million input tokens
- Recall: $0.75 per million output tokens
- Reflect: $0.05 per call
- Retrieve Model: $0.25 per million output tokens
- Refresh Model: $0.05 per call
- File Conversion: $7.50 per million output tokens
- Storage: $0.25 per million tokens per month, free for the first 30 days after a memory is retained

## Can Codex, Claude Code and Hermes share one Hindsight memory?

Yes, through MCP. Hindsight exposes an MCP endpoint per memory bank, enabled by default at `http://localhost:8888/mcp/{bank_id}/`. The repository names Claude Code, Cursor, Codex and GitHub Copilot as clients, and xAI's Grok Build CLI also supports MCP.

There are two caveats. An endpoint on a private Tailscale network is reachable by local clients on that network, not by a cloud-hosted assistant, so "even Grok" depends on which Grok client you mean. The video also shows vendor logos rather than four agents reading one bank, and the banks visible on screen are named "atlas" and "hermes-agent," which suggests separate banks.

The video ends by asking viewers to comment "Memory" for a guide and shows a page promoting the creator's paid Agent Operator Skool community.

## Bottom line

Self-hosted Hindsight is free software that gives Hermes and other MCP clients a shared, uncapped store of raw memories. It still consolidates what the model sees, its Cloud version bills per operation, and it supplements Hermes' 2,200 and 1,375 character memory files instead of replacing them.

## Sources

- [Hindsight overview](https://hindsight.vectorize.io/) - operations and memory types
- [vectorize-io/hindsight on GitHub](https://github.com/vectorize-io/hindsight) - license, deployment options, MCP endpoint and providers
- [The Consolidation Problem in Agent Memory](https://hindsight.vectorize.io/blog/2026/05/21/agent-memory-consolidation) - Hindsight's consolidation policy
- [Hindsight Cloud billing](https://docs.hindsight.vectorize.io/billing/) - Cloud pricing
- [Hermes Agent persistent memory docs](https://hermes-agent.nousresearch.com/docs/user-guide/features/memory) - character limits and overflow behaviour
- [Hermes Agent memory providers docs](https://hermes-agent.nousresearch.com/docs/user-guide/features/memory-providers) - external provider rules
- [Hindsight integration guide for Hermes](https://hindsight.vectorize.io/sdks/integrations/hermes) - setup and built-in memory mirroring
- [Hermes Agent homepage](https://hermes-agent.nousresearch.com/) - Nous Research's agent


## Video description and verification notes

**A 36.288-second vertical clip (0:36), 1080x1920 at 30 fps, HEVC Main video at roughly 492 kbps alongside an HE-AAC stereo track at 44.1 kHz and roughly 96 kbps, 2,697,142 bytes on disk (2.6 MB), posted 2026-10-05 at 22:14:55 UTC by @josephbchandler under the channel nickname "Joseph Chandler." At capture on 2026-10-06 it showed 4,838 views, 139 likes, 9 comments, and 19 reposts. The audio is listed as "original sound" credited to Joseph Chandler, which is the speaker's own voice recorded on a clip-on lavalier he holds in frame with its green power LED lit. I read all 18 extracted frames and a 117-word transcript. Frame 1 is an Apple Mac mini product page behind the speaker, with the partially cropped marketing copy "...ing feel extra / ...ime on Mac mini, / ...U speed up all kinds." Frame 2 cuts to a saturated blue site titled "HERMES AGENT" with an "INSTALL HERMES" button, a hamburger menu, the headline "THE AGENT THAT G... WITH Y...", a "DOWNLOAD FOR MACO[S]" button with an Apple logo, the label "INSTALL VIA TERMINAL," and a tabbed code box offering "macOS" and "[W]indows" with a copy icon. Frames 4 and 5 show a dense force-directed graph of pink and purple nodes on white, with legible memory strings including "The user requires clearly iden...", "The assistant reported verifyi...", "The assistant reported repa...", and date-stamped nodes reading "...ober 4, 2026, backs ground..." and "On October 4, 20...". Frames 7 and 8 show a desktop window titled "Persistent memory" with a search field reading "Search ⌘ K" and the breadcrumb "Settings > Memory & Context > Persistent memory." Inside it: "Persistent Memory / Save durable memories that can help future sessions" with a blue toggle switched on, "User Profile / Maintain a compact profile of..." also toggled on, a "Memory Budget" field containing 2200, a "Profile Budget" field containing 1375, and a "Memory Provider / Memory provider plugin" dropdown whose value reads "Built-i" before the speaker's shoulder covers it. Frame 9 is the Hindsight marketing page on black: "Agent Memory That Learns," the subhead "State of the art long-term memory for your agents.", a gradient "Start free on Cloud →" button, a star counter reading "45" with its remaining characters occluded, and a partially hidden line reading "$5 of free cr...". Below it sits a product screenshot with a bank selector reading "hermes-agent," a "GitHub" link, a sun icon, a "Color by: Mentioned" control, "Link types," a "Pause" button, and a "Fullscreen" button over date labels "2026-09-30" and "2026-09-30." Frames 10 through 12 are black motion-graphic cards showing a speech-bubble icon over scattered dots, then three dots joined into a triangle, then a robot-head icon wired down to a single lit node. Frames 13 and 14 return to a browser at the address `atlas.tail49ce64.ts.net` showing the Hindsight app with a bank picker reading "atlas," an "+ Add Document" button, the page title "Home / Overview of shared-personal", a "Memory constellation" panel with "View all →", "Scroll to zoom," "Drag to pan," "Hover to explore," "Share," "Pause," "Fullscreen," a LINKS legend running "few" to "many," a counter reading "200 memorie[s]", node labels "The assistant reported verifyi...", "On October 5, 2026, ...", "...ermes Age...", a hash fragment "cf7f9b92f418a7", and a "Knowledg[e]" panel below. Frames 15 through 17 add vendor marks one at a time on black: the OpenAI mark on "so my Codex," a sunburst mark and an anime-style avatar icon on "Hermes can," and in frame 18 a white rounded face icon on "even Grok," all four wired to a central node. The final frame is a browser at `file.joseph.so` showing a page branded "Hypertools" with an "Agent Operator ↗" link, body text "which setup makes sense for you.", the byline "By Joseph Chandler / Hypertools", "Checked October 5, 2026", a dark callout labeled "24/7 AI AGENT OPERATION" reading "Want to master 24/7 AI age[nts]?" and "Join my Agent Operator Sko[ol]" with the line "Price increases at launch" and a "Get free access now ↗" button, then a nav row reading "Quick picks / Compare / Obsidian" and a section header "01 / START HERE / Quick picks."**

## The "no limit" claim contradicts Hindsight's own published consolidation policy

The clip draws a clean line: "Hermes does come with its own built-in memory, but as it fills up, the agent has to condense or remove things. Hindsight is better. There's no limit."

Hindsight's own engineering blog rejects that framing. [The Consolidation Problem in Agent Memory](https://hindsight.vectorize.io/blog/2026/05/21/agent-memory-consolidation) opens with "An agent that remembers everything is an agent that remembers nothing useful," and then specifies that "Every consolidation policy operates on four levers: 1. Importance — Which observations become memories at all. 2. Merge — How related facts get unified into a single canonical record. 3. Decay — How confidence in a memory degrades over time. 4. Eviction — When a memory leaves the system entirely."

Hindsight runs all four. The [Hindsight overview](https://hindsight.vectorize.io/) describes a tiered structure of world facts, experience facts, observations, mental models, and knowledge pages, where observations are consolidated beliefs derived from raw facts and mental models are curated summaries derived from observations. That is condensing, performed on a schedule instead of under duress.

One part of the claim survives. Hindsight treats eviction differently from Hermes: "eviction is a compliance tool, not a performance tool. For GDPR, user-requested deletion, and PII redaction, it is non-negotiable." Hindsight does not delete raw facts to free capacity. Hermes does. So raw storage is effectively unbounded while the material an agent actually reads back is compressed and decayed on purpose. "There's no limit" is accurate about the store and misleading about what reaches the model.

## "It's free" holds for the MIT self-hosted build and not for the button on screen

The speaker says "For that, I'm running Hindsight. It's free," over the page in frame 9. That same frame shows a "Start free on Cloud →" button and a partially occluded "$5 of free cr..." line. Hindsight Cloud is metered. The [Hindsight Cloud billing page](https://docs.hindsight.vectorize.io/billing/) lists Retain at $10.00 per million input tokens, Recall at $0.75 per million output tokens, Reflect at $0.05 per call, Retrieve Model at $0.25 per million output tokens, Refresh Model at $0.05 per call, File Conversion at $7.50 per million output tokens, and Storage at $0.25 per million tokens per month, with the note that "Memories stored in your banks are free for the first 30 days after they're retained." The $5 signup credit is a trial allowance, not a free tier.

The speaker's own setup resolves this. Frames 13 and 14 show his Hindsight instance at `atlas.tail49ce64.ts.net`, a Tailscale tailnet hostname, which means he is self-hosting rather than using Cloud. The [Hindsight repository](https://github.com/vectorize-io/hindsight) is MIT licensed and ships a Docker image, a pip install, a Helm chart, and a Python embedded mode, so the software itself costs nothing. Inference still does. Hindsight calls a model on every retain and every reflect, and the project documents `HINDSIGHT_API_LLM_PROVIDER=ollama` as the route to avoid API charges entirely by running the extraction model locally. On a Mac mini that is a plausible setup, but the video never says which provider he is paying for, so the true running cost is unverified.

## Hindsight supplements Hermes built-in memory, it does not replace it

The implied upgrade path in the clip is that Hindsight fixes the condense-or-remove problem. The Hermes integration docs say otherwise. Hindsight's [Hermes integration page](https://hindsight.vectorize.io/sdks/integrations/hermes) states that with both active "the model may prefer the built-in one," and recommends disabling MEMORY.md and USER.md if you want Hindsight to be the primary store. When both stay enabled, "every entry the agent adds or replaces in them is also saved to the bank, tagged `builtin-memory`." The [Hermes memory providers documentation](https://hermes-agent.nousresearch.com/docs/user-guide/features/memory-providers) adds that external providers run alongside built-in memory and that only one external provider can be active at a time.

So installing Hindsight does not lift the Hermes budget ceiling. It mirrors writes out of the capped files into an uncapped bank, which preserves what would otherwise be pruned, while the capped files keep getting pruned. Frame 8 reinforces that the switch had not been thrown at capture: the "Memory Provider" dropdown reads "Built-i" before the speaker's shoulder covers it.

## The 2200 and 1375 on the settings screen are character limits, not token budgets

The Hermes UI labels these fields "Memory Budget" and "Profile Budget," which reads as a token allowance. The [Hermes persistent memory docs](https://hermes-agent.nousresearch.com/docs/user-guide/features/memory) specify characters: `memory_char_limit: 2200`, roughly 800 tokens, and `user_char_limit: 1375`, roughly 500 tokens. Both files live under `~/.hermes/memories/` and are injected into the system prompt as a snapshot at session start, and the combined footprint is kept small on purpose so the prefix stays stable for caching.

The video's description of the overflow behavior is close to right but imprecise. Hermes does not auto-trim in the background. The docs say the write tool "returns an error instead of silently dropping entries," after which the agent "makes room itself — consolidating or removing entries in the same turn before retrying." The agent spends a turn on cleanup rather than quietly losing data. The numbers in frames 7 and 8 are the stock defaults, so the configuration shown is unmodified.

## Cross-agent sharing is real over MCP, and the end card is the actual product

"Hindsight memory works across all my agents, so my codex and my cloud code and my Hermes can all share the same memories, even GrockBot." Whisper mangled two names: the on-screen captions read "Codex," "Hermes," and "Grok," and "cloud code" is Claude Code. The mechanism checks out. Hindsight ships a built-in MCP endpoint per bank, enabled by default at `http://localhost:8888/mcp/{bank_id}/`, and the repository names Claude Code, Cursor, Codex, and GitHub Copilot as clients. xAI's Grok Build CLI also speaks MCP. Pointing several clients at one bank is a supported configuration.

Two caveats the clip skips. A tailnet-hosted endpoint like `atlas.tail49ce64.ts.net` is reachable by local MCP clients on the same Tailscale network, not by a cloud-side assistant fetching a remote MCP server, so "even Grok" depends on which Grok client is involved. And the four icons in frames 15 through 18 are vendor marks only, not a demonstration of four agents reading the same bank. The screen shows one bank named "atlas" with "200 memorie[s]" and a bank selector reading "hermes-agent" in a separate shot, which is consistent with multiple banks rather than one shared pool.

The closing frame shows what the comment gate actually leads to. The page at `file.joseph.so` is branded "Hypertools," bylined "By Joseph Chandler / Hypertools," dated "Checked October 5, 2026," and carries a callout reading "Want to master 24/7 AI age[nts]?", "Join my Agent Operator Sko[ol]", "Price increases at launch", and "Get free access now ↗". The video's description field is the same seven words as its title, "Comment Memory for the guide," with no link. Nine comments had landed at capture against 4,838 views. The delivery method for the promised guide is never stated on screen or in the description.

## Key Takeaways

- **Verified:** Hindsight is MIT licensed, built by Vectorize, and ships Docker, pip, Helm, and Python embedded deployment paths, so the self-hosted build on the tailnet address shown costs nothing in license fees.
- **Verified:** Hindsight exposes an MCP endpoint per bank by default, and Claude Code, Codex, Cursor, GitHub Copilot, and xAI's Grok Build CLI are all MCP clients, so pointing several agents at one bank is supported.
- **Correction:** "There's no limit" conflicts with Hindsight's published consolidation policy, which runs importance, merge, decay, and eviction levers and compresses raw facts into observations and mental models.
- **Correction:** Hindsight does not replace Hermes built-in memory. Only one external provider can be active at a time, it runs alongside MEMORY.md and USER.md, and the docs warn that "the model may prefer the built-in one" unless you disable them.
- **Partial correction:** "It's free" is true of the self-hosted MIT build and false of the "Start free on Cloud" button on screen in the same shot, which bills per operation after a $5 signup credit.
- **Partial correction:** Hindsight does not evict for capacity, so raw storage really is unbounded. What reaches the model is still compressed and decayed by design.
- **Context:** The 2200 and 1375 shown in the Hermes settings panel are character limits, not token budgets, and they are the stock defaults.
- **Unverified:** Which LLM provider backs the speaker's Hindsight instance, and therefore his actual running cost. I checked the video, the description, and the visible UI, and none of them name it.
- **Unverified:** How the "Comment Memory" guide is delivered. No link appears in the description, and nothing on screen states a delivery method.

## Resources

- [Hindsight overview](https://hindsight.vectorize.io/). Establishes Hindsight as a Vectorize product, defines retain, recall, and reflect, and lists the five memory types including the consolidated observation and mental model tiers.
- [vectorize-io/hindsight on GitHub](https://github.com/vectorize-io/hindsight). Confirms the MIT license, the Docker and Helm and embedded deployment paths, the built-in per-bank MCP endpoint, and support for 25+ LLM providers including local Ollama.
- [The Consolidation Problem in Agent Memory](https://hindsight.vectorize.io/blog/2026/05/21/agent-memory-consolidation). Vectorize's own statement of the four consolidation levers and the rule that eviction is a compliance tool rather than a capacity tool.
- [Hindsight Cloud billing](https://docs.hindsight.vectorize.io/billing/). Lists the per-operation rates for retain, recall, reflect, model retrieval and refresh, file conversion, and storage, plus the 30-day storage grace period.
- [Hermes Agent persistent memory docs](https://hermes-agent.nousresearch.com/docs/user-guide/features/memory). Specifies memory_char_limit 2200 and user_char_limit 1375 as character limits and describes the error-and-retry overflow behavior shown in the settings panel.
- [Hermes Agent memory providers docs](https://hermes-agent.nousresearch.com/docs/user-guide/features/memory-providers). States that external providers run alongside built-in memory and that only one external provider can be active at a time.
- [Hindsight integration guide for Hermes](https://hindsight.vectorize.io/sdks/integrations/hermes). Gives the setup command and config keys, and documents the builtin-memory mirroring and the warning that the model may prefer built-in memory.
- [Hermes Agent homepage](https://hermes-agent.nousresearch.com/). Confirms the Nous Research attribution, the MIT license, and the "The Agent That Grows With You" headline partially visible in frame 2.
