Persistent memory without flipping a config flag

Hermes Agent's most interesting feature is the closed learning loop: skills the agent generates from its own experience that compound over months of use. The published TokenMix benchmarks show self-created skills cutting research-task time by 40% versus a fresh agent.

Hermes Agent's most common complaint is that the same feature is off by default. Reviewers downloaded v0.10.0, ran it with stock config, and concluded the agent forgot every conversation. Many wrote dismissive reviews calling it "nothing special." They had never enabled persistent memory or skill_generation in the config; the version of Hermes most users run is structurally not the version of Hermes the product is supposed to be.

This post is about the alternative contract: memory and learning on by default, no config flags, value unlocked from minute one.

Why off-by-default is structurally a problem

Three reasons memory tends to ship off-by-default in self-hosted agents:

Resource cost. Memory storage and consolidation has real CPU and storage cost. Engineers running on personal infrastructure want the choice to opt in.

Privacy ergonomics. Persistent memory means the agent has a long log of your conversations on disk. Some users genuinely do not want this; off-by-default lets them choose.

Behavioral safety. A misbehaving agent with persistent memory can carry mistakes across sessions. Off-by-default isolates the blast radius of a bad day.

These are reasonable engineering choices. They are also choices that mean the average user (who does not flip the flag) does not get the feature.

For non-engineers, the right contract is the opposite: on by default, opt out if you specifically want to. The reasoning: the value of the feature is high, the resource cost is dwarfed by the user-experience benefit, and the privacy/safety concerns are addressable through other means (per-user isolation, audit logs, granular forget controls).

What "on by default" looks like operationally

For Ajeris specifically, memory is not one feature; it is several layers all running from minute one:

• Working memory. The current conversation's context. Always on; that part is universal.
• Episodic memory. What was said and done across past conversations. Embeddings indexed; recallable by semantic search. On from minute one.
• Semantic memory. Stable facts about the user and their world. Built up from episodic memory through consolidation.
• Core memory. The small set of facts that should never be forgotten. The user's name, the people in their life, the routines they want respected.
• Skills as procedural memory. The agent saves how-to procedures from successful interactions. Auto-versioned with timesUsed and lastUsedAt tracking.

All five layers run by default. There is no flag to flip. There is no config to read. The agent remembers from minute one because the contract is "remember unless told to forget."

Consolidation: the night work

The other layer most self-hosted agents leave off by default is the consolidation job. Memory that just accumulates without consolidation grows linearly with use; eventually the relevant facts are buried under situational noise that no one needs.

Consolidation is the nightly job that promotes important memories into long-term storage and lets the rest fade. For Ajeris:

• Dreaming worker. 3am UTC, fleet-wide. Fans out to per-user consolidation. Identifies facts worth promoting.
• Daily reflection. 11pm UTC, per-user. Writes a 2-3 sentence summary of the day. May save a skill if the user solved something interesting.
• Memory eviction. When a memory has not been recalled in months and is structurally redundant, it can be evicted to keep the working set sharp.

These run automatically. The user does not configure them. They simply happen overnight and the agent is sharper the next morning.

What this changes about daily use

Three concrete shifts:

The agent stops asking you to repeat yourself. Every conversation does not start with "remind me who you are again." The agent already knows.

The agent stops being situationally useful only. A chat product helps you with the question you are asking right now and forgets it tomorrow. A memory-on agent helps you with the cumulative pattern of your life: it remembers your spouse's name, your kid's allergies, the contractor you used last year, the routine you set up for Tuesdays.

The agent gets better over time. Skills accumulate. The agent that helped you book the dentist last month knows how you like the appointment scheduled this month. The pattern compounds.

Why Hermes Agent is the right product for the right user

If you are an engineer who wants the open-source self-improving agent, who is willing to read the docs, flip the right flags, and enable the closed learning loop, Hermes Agent is genuinely strong. The 40% research-task speedup benchmark is real. The architecture is sound. We are not arguing the product is bad.

We are arguing that for the much larger population of users who will not flip flags, the contract has to be on-by-default. That is the case for the hosted product.

Try it

Free to start. Text +1 (470) 821 4265. The test is using the agent for a week and noticing what it remembers about you on day 7 that you never told it twice.

For the head-to-head comparison, see Ajeris vs Hermes Agent.