L3 Persona Tier

Status: partially wired. Read What is not wired before relying on this.

The persona tier holds a small set of durable behavioural facts about the user. It is a store, not a prompt channel: promotion writes a row to the memoria_persona table, while prompt injection reads the opt-in persona.md file and never queries the table. Promoting a fact changes nothing about your prompts until something regenerates that file.

The intended role is a set of facts that are always present rather than competing for recall, and that role has a cost: anything that reaches the injected file consumes prompt budget on every turn, so the tier is meant to stay small.

---

Storage

Table memoria_persona, created in init_beam (mnemosyne/core/beam.py:988):

Column	Type
id	INTEGER PRIMARY KEY AUTOINCREMENT
session_id	TEXT DEFAULT 'default'
tier	TEXT NOT NULL CHECK(tier IN ('permanent','long_term','working'))
topic	TEXT NOT NULL
content	TEXT NOT NULL
confidence	REAL DEFAULT 0.7
source_memory_id	TEXT (soft reference, no FK)
created_at	TIMESTAMP DEFAULT CURRENT_TIMESTAMP
last_reinforced_at	TIMESTAMP DEFAULT CURRENT_TIMESTAMP
reinforcement_count	INTEGER DEFAULT 0
promotion_reason	TEXT

Indexed on (session_id, tier) and (tier, topic).

Tiers

Three values are accepted: permanent, long_term (the default on promote), and working.

Tier affects mnemosyne_persona_list ordering only. Listing sorts permanent before long_term before working, then by reinforcement_count descending. Nothing else consults the tier: the prompt path reads persona.md and never queries this table.

Tier does not cause retention or expiry. There is no code anywhere that deletes, ages out, or reduces the confidence of a memoria_persona row. The only writes are insert on promote, delete on demote, and the reinforcement counter update. Descriptions elsewhere in the codebase that call permanent "never evicted" or long_term "reinforcement-driven decay" describe an intended design, not current behaviour: nothing is evicted, because no eviction exists.

Operations

Implemented by PersonaAdapter in hermes_memory_provider/persona_adapter.py, duplicated at integrations/hermes/src/mnemosyne_hermes/persona_adapter.py. All methods return JSON strings and never raise.

Operation	Behaviour
promote	Reads content from working_memory, falling back to episodic_memory. Derives topic from memoria_timelines via source_memory_id, else "general". Errors if the content is empty. Default tier long_term.
demote	Not a plain delete. Inserts a tombstone into memoria_preferences recording [demoted from <tier>] <content> and the reason, then deletes the memoria_persona row. Both in one transaction.
list	Optional tier and topic filters. Returns rows in tier order (permanent first), then by reinforcement_count descending.
reinforce	reinforcement_count += 1 and last_reinforced_at = CURRENT_TIMESTAMP. Returns an error if the id does not exist.

Reinforcement feeds two consumers: the list ordering above, and the regeneration trigger, which uses MAX(last_reinforced_at) as its watermark.

MCP tools

mnemosyne_persona_promote, mnemosyne_persona_demote, mnemosyne_persona_list, mnemosyne_persona_reinforce.

These four are advertised over MCP but are not callable over MCP. Their schemas are in ALL_TOOL_SCHEMAS, so an MCP client sees them in list_tools, but there is no entry in _TOOL_HANDLERS, so calling one returns Unknown tool. They work only through the Hermes plugin, which routes any mnemosyne_persona_* prefix to the adapter.

There is no CLI surface for any of them.

Prompt injection

HermesPersonaPromptMixin in mnemosyne/integrations/hermes_persona_prompt.py appends a block headed # L3 Persona (Active Behavioral Rules) to the system prompt.

The injected content comes from a markdown file on disk, not from the memoria_persona table. The file is read at prompt-build time and cached on its mtime, so a warm prompt does no file IO. When persona is disabled the function returns early with no file access at all. Any error reading the file degrades to an empty block.

Injection applies a word-based cap: token_cap * 0.75 words. On overflow it truncates and then trims back to the last ## section boundary so a section is never cut mid-way, appending a note that the content was truncated.

Configuration

All five variables are read with os.environ.get at module import time. Setting them afterwards has no effect.

Variable	Default	Controls
MNEMOSYNE_PERSONA_ENABLED	false	Master switch for prompt injection
MNEMOSYNE_PERSONA_FILE	~/.hermes/memory/persona.md	The markdown file written and injected
MNEMOSYNE_PERSONA_TOKEN_CAP	1500	Prompt budget for the injected block
MNEMOSYNE_PERSONA_INTERVAL	50	New working memories before the regeneration trigger fires
MNEMOSYNE_PERSONA_DAILY_SYNC_HOUR	3	UTC hour for the daily trigger. Any value outside 0-23 disables it

config.yaml does not reach the persona code. Four of these keys exist in ENV_VAR_MAP, but neither the persona module nor the prompt mixin consults MnemosyneConfig, so only the environment variables take effect. The defaults declared in config.py also disagree with the real ones above (it declares persona_enabled: true, persona_token_cap: 500, persona_interval: 10). Trust this table and the environment. MNEMOSYNE_PERSONA_FILE has no config key at all.

What is not wired

Being explicit, because the surrounding docstrings are optimistic:

1. Nothing calls mnemosyne/core/persona.py. PersonaExtractor, PersonaTriggers, render_persona_markdown, and write_persona_file have no callers outside the test suite. There is no scheduler, no sleep hook, and no CLI command that regenerates persona.md. The five trigger conditions and the daily-sync hour are library plumbing that a caller must drive.
2. The file the injector reads must be produced by you or by your own code calling write_persona_file. Promoting a fact into memoria_persona does not update persona.md, so promotion alone changes nothing about your prompts.
3. No decay or eviction, as described under Tiers.
4. The render-side and injection-side token caps use different algorithms (word count times 1.3 versus a 0.75 word multiplier) and disagree slightly on what fits.

If you want the tier working end to end today, the shape is: promote facts through the Hermes tools, then call write_persona_file yourself on whatever schedule you want, with MNEMOSYNE_PERSONA_ENABLED=1 and MNEMOSYNE_PERSONA_FILE pointing at the same path.

See also

• Architecture for where the tier sits relative to BEAM
• Generated configuration reference
• docs/roadmap-layered-agent-memory.md for the L0-L4 model this tier is named after